Models

Qwen3.8 27B is a dense, open-weight multimodal model from Alibaba's Qwen team, released August 14, 2026 as a smaller member of the Qwen3.8 family alongside the flagship Qwen3.8 Max. It combines Gated DeltaNet linear attention with standard gated attention across 64 layers, giving a 27 billion parameter dense model a native 262K token context window, extendable to 1M tokens. It accepts text, image, and video input, including hour-scale video and STEM diagrams. Alibaba reports 61.7 on SWE-bench Pro and 73.0 on Terminal Bench 2.1, both improvements over the earlier Qwen3.6 27B, and 89.2 on GPQA Diamond. Released under Apache 2.0, it gives developers an open-weight alternative to Qwen3.8 Max for coding and agentic tasks, at a fraction of the parameter count.

by Qwen | $0.45/M input | $3.2/M output

Dots3-Note Preview is a multimodal mixture-of-experts model from Dots Studio, the AI lab at Xiaohongshu (RedNote), also behind the dots.llm and dots.ocr models released under the rednote-hilab name. It has 280 billion total parameters, 16 billion active per token, and accepts text, image, video, and audio input. The model introduces TEMPO, a reinforcement learning method for long-horizon agent tasks. It periodically checks its own progress and updates its memory while working, targeting agents that operate over hours. Dots Studio reports 78.4% on SWE-bench Verified and 79.1% on MMMU Pro, and says a checkpoint scored a perfect 42/42 on the 2026 International Mathematical Olympiad, matching Huawei's Celia and beating Gemini Deep Think's 35/42 at the same event. These are vendor-reported figures, not yet independently verified. This free-tier listing is served through OpenRouter, with a 512K-token context window suited to long documents, codebases, and multi-step agent workflows.

by Dots Studio | $0/M input | $0/M output

Gemini 3.7 Flash is Google's workhorse Flash-tier model, released August 13, 2026, three weeks after Gemini 3.6 Flash. It's built for coding and agentic workflows, targeting software engineering, web development, and knowledge-dense domains like finance and law. Google reports gains over Gemini 3.6 Flash on several benchmarks. DeepSWE v1.1 rose from 49.0% to 65.3%, FrontierCode 1.1 from 34.4% to 43.6%, and AutomationBench from 17.0% to 30.4%. On FrontierCode 1.1 it scores above Claude Sonnet 5 (42.7%) and GPT-5.6 Terra (41.3%), though GPT-5.6 Terra edges it out on Terminal-bench 2.1 (87.4% vs 85.8%). It accepts text, image, video, audio, and PDF input with a 1M token context window and 64K token output limit. It supports function calling, search as a tool, and computer use, and has a March 2026 knowledge cutoff. It's priced at roughly half of Gemini 3.6 Flash's rate, fitting teams running coding agents or high-volume document processing.

by Google | $0.38/M input | $1.88/M output

Grok 4.6 is xAI's frontier model for long-running agents and interactive, visual development work, built on the same 1.5-trillion-parameter V9 foundation as Grok 4.5 with additional supervised fine-tuning and reinforcement learning. It matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index (both score 61) and leads on CursorBench v3.2 at 69.9%, with gains over Grok 4.5 across coding, research, and agentic benchmarks. xAI highlights it for vulnerability patching, engineering design work, and turning a rough product idea into a working first version, with improved self-testing on longer tasks. The model accepts text and image input with text-only output, a 500,000 token context window, and no output token limit. Reasoning effort is configurable across four levels (low, medium, high, xhigh), and it supports function calling, web search, X search, and code execution. At $2 per million input tokens and $6 per million output tokens, it costs the same as Grok 4.5.

by xAI | $2/M input | $6/M output

Qwen3.8 2.4T A95B is Alibaba's open-weight release of its Qwen3.8 Max flagship, a sparse mixture-of-experts model with 2.4 trillion total parameters and 95 billion active per token, routed across 512 experts. It uses a hybrid attention design (Gated DeltaNet and Gated Attention layers) across 92 layers, with a native 262K context window and thinking mode enabled for every response. Alibaba reports 93.0 on PaperBench (ahead of GPT-5.6 Sol's 90.5), 92.6 on GPQA Diamond, 86.6 on Terminal-Bench 2.1, and 67.7 on SWE-bench Pro, positioning it for coding, research, and long-horizon agentic work. It gives developers access to Qwen-Max-class capability under open weights, useful for teams that want frontier-level coding and agentic performance without a closed API.

by Qwen | $1.8/M input | $5.4/M output

DeepSeek V4 Pro 0813 is the production release of DeepSeek's V4 Pro flagship model, a 1.6-trillion-parameter Mixture-of-Experts model with 49B parameters active per token, marking the end of a four-month preview. It keeps the preview's 1,048,576-token context window, 384,000-token max output, and three reasoning modes (non-thinking, high, and max effort). DeepSeek reports large gains over the preview build: Terminal Bench 2.1 rose from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7. At max reasoning effort it scores 80.6% on SWE-bench Verified, matching Gemini 3.1 Pro, along with 90.1% on GPQA Diamond, 87.5% on MMLU-Pro, and a Codeforces rating of 3,206, ahead of GPT-5.4's 3,168. Priced at $0.64 per million input tokens and $1.91 per million output tokens. Pinning the 0813 route locks an application to this exact checkpoint rather than the rolling deepseek-v4-pro alias, useful for agentic coding and long-context reasoning tasks.

by DeepSeek | $0.64/M input | $1.91/M output

Seed 2.1 Turbo is ByteDance's faster, lower-cost variant of the Seed 2.1 model family, released alongside Seed 2.1 Pro at its Volcano Engine FORCE conference in June 2026. It carries the same feature set as Pro, including coding engineering across full project cycles, long-chain agent workflows with multi-step tool use, and multimodal understanding of text, image, and video. ByteDance says the Seed 2.1 series matches GPT-5.5 on these three capability pillars, though that claim is vendor-reported and has not been independently verified. No Turbo-specific benchmark scores have been published separately from Pro. Turbo is priced at about half of Seed 2.1 Pro per token, which suits enterprise teams running high volumes of API calls, such as agent pipelines and document or image processing, where cost and latency matter more than peak capability.

by ByteDance Seed | $0.5/M input | $2.5/M output

Sakana Namazu is a Japanese-specialized large language model from Tokyo-based Sakana AI, built on Moonshot AI's Kimi K2.6 and further tuned on Sakana's own data. The tuning targets the nuances of Japanese language and business practice, including keigo honorifics, business customs, and everyday slang that general-purpose models often miss. Sakana AI reports the tuning also improves neutrality and factual accuracy on topics specific to Japan, while keeping reasoning, knowledge, and coding performance in line with the base model. The API is OpenAI-compatible and ships with built-in web search and code execution tools, so it can be dropped into existing agent and chat integrations with minimal changes. Sakana Namazu suits developers building Japanese-language products, such as customer support, business writing, and research or coding agents that need to work naturally in Japanese.

by Sakana AI | $0.95/M input | $4/M output

Nemotron 3.5 Lightning is a 30B-parameter Mixture-of-Experts model from NVIDIA with 3B active parameters per token, using a hybrid Mamba-2, MoE, and attention layer architecture. It is designed for high-volume, low-latency execution inside multi-agent systems, where a larger reasoning model like Nemotron 3 Ultra plans and delegates, and Lightning handles tool calls, code review, and other repetitive subagent tasks. NVIDIA reports up to 4x higher output throughput than similarly sized open models and, on its PinchBench agent benchmark, 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy. The model ships with built-in multi-token prediction and draft models for speculative decoding, and is available in NVFP4 and BF16 checkpoints under the permissive OpenMDW-1.1 license with open weights and training recipes. Choose it for agent harnesses handling frequent, narrow calls, tool validation, or subagent delegation where response speed matters more than broad reasoning depth.

by NVIDIA | $0.08/M input | $0.2/M output

Solar Pro 4 is an agentic large language model from Upstage, built for multi-step work that spans documents, terminals, and tool calls, with a 524K token context window. Compared to Solar Pro 3, it shows large gains on agent-focused evaluations. Terminal-Bench v2.1 rose from 43.2 to 57, and AA-LCR long-document reasoning rose from 62.7 to 71. It scores 89.0 on GPQA Diamond and 81.9 on Arena-Hard v2 for Japanese. The model supports English, Korean, and Japanese, with adjustable reasoning effort for either deep analysis or faster responses. It is designed to flag when information can't be verified rather than fabricate an answer. Solar Pro 4 targets contract review, multi-document reconciliation, terminal-based automation, and document-heavy workflows like generating Excel, Word, and PowerPoint outputs.

by Upstage AI | $0.03/M input | $0.12/M output

Muse Glimmer 30B is a 30-billion-parameter dense model from Meta Superintelligence Labs, pairing a causal transformer with a 1.8-billion-parameter vision encoder for text and image input. It's released under an Apache 2.0 license, Meta's first fully open-weight model since Muse Spark moved to a paid API. On Meta's own benchmarks, it scores 76.0 on SWE-Bench Verified, 51.2 on SWE-Bench Pro, 94.7 on AIME 2026, 83.5 on GPQA Diamond, 75.5 on MCP Atlas, and 74.6 on DeepSearch QA, ahead of similarly sized open models like Gemma4 31B and Qwen3.6 27B on Meta's reporting. These figures are vendor-reported and not independently verified. Where Muse Spark targets large-scale multi-agent orchestration, Muse Glimmer sits a size tier down, aimed at tool use, multi-step reasoning, coding, and LLM-as-a-judge evaluation with a 131,072-token context window. It fits developers who want agentic and coding capability at lower cost than the larger Muse Spark models.

by Meta Llama | $0.35/M input | $1.5/M output

Muse Spark 1.2 is Meta Superintelligence Labs' coding-focused update to Muse Spark 1.1, released alongside Muse Code, a terminal coding agent it powers. Meta scaled up training compute on coding tasks and training-environment diversity, aiming at code generation, debugging, codebase understanding, and long-horizon work like whole-repository generation. On Meta's own evaluation harness, Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1, edging OpenAI's GPT-5.6 Terra (81.8%) and xAI's Grok 4.5 (81.6%) on Terminal-Bench but trailing Anthropic's Opus 5 (86.7%). These are vendor-run numbers, not yet independently verified on the public leaderboards. It keeps the 1,048,576-token context window and text, image, video, audio, and PDF input from 1.1, at the same $1.25 per million input and $4.25 per million output token pricing. A separate contributor tier offers lower rates in exchange for letting Meta train on your prompts and completions.

by Meta Llama | $1.25/M input | $4.25/M output

Qwen3.8 Max is Alibaba's flagship large language model, released August 3, 2026 as the most capable model in the Qwen family to date. It uses a mixture-of-experts architecture with 2.4 trillion total parameters and about 95 billion active per request, and accepts text, image, and video input with a context window of up to 1 million tokens. Alibaba positions it for coding and long-horizon agentic work: in testing the model ran autonomously for over 10 days building a self-evolving software harness. Reported benchmarks include 93.0 on PaperBench, 82.8 on IFBench, 86.6 on Terminal-Bench 2.1, and 86.1 on OSWorld-Verified, ahead of Claude Opus 4.8 on several coding and agent tasks and roughly matching Claude Fable 5 and GPT-5.6 Sol, though it trails both on some evaluations. On the Arena.AI leaderboard it ranks as the top Chinese model for text tasks. Alibaba plans to open-source the weights on Hugging Face and ModelScope.

by Qwen | $0.83/M input | $2.48/M output

DeepSeek V4 Flash 0731 is the July 31, 2026 dated snapshot of DeepSeek's V4 Flash, a sparse Mixture-of-Experts chat model with a 1,048,576-token context window, hosted here on Together AI's infrastructure. It keeps the April preview's architecture but was re-post-trained with a pipeline focused on coding, reasoning, agents, and tool use. In DeepSeek's own benchmarks the agentic gains over the preview are large: Terminal Bench 2.1 rises from 61.8 to 82.7, Cybergym from 38.7 to 76.7, and DeepSWE from 7.3 to 54.4. On these vendor-reported numbers it also outscores the V4 Pro preview, though third-party reproduction was not available at release. Pinning the dated 0731 route locks an application to this exact checkpoint, while a rolling deepseek-v4-flash route can move to newer versions. It fits high-volume coding assistants, chat systems, and agent workflows where cost and latency matter, and this route's 1M-token max output gives more headroom for long agent traces than the original preview route.

by DeepSeek | $0.14/M input | $0.28/M output

Qwen3.7 Flash is the low-cost, fast tier of Alibaba's Qwen3.7 family, released in July 2026. It is a vision-language model that accepts text, image, and video input across a 1 million-token context window, with reasoning enabled by default and a 262K-token thinking budget. Alibaba positions it as an upgrade over Qwen3.6 Flash in multimodal understanding and agent execution, with better object recognition and spatial intelligence. It supports function calling and structured outputs. Pricing is tiered by prompt length. Requests under 32K input tokens cost $0.03/$0.13 per million, rising to $0.20/$0.80 above 256K. Alibaba published no benchmarks at launch. An independent vision evaluation by Roboflow measured strong object identification (84.4%) but weak OCR and object detection, so it fits high-volume multimodal tasks (classification, visual agents, lightweight extraction) better than document-heavy pipelines.

by Qwen | $0.03/M input | $0.11/M output

Claude Opus 5 is Anthropic's flagship model, released July 24, 2026 as the successor to Claude Opus 4.8. Anthropic positions it as approaching the performance of its higher-tier Claude Fable 5 model at roughly half the token cost, built for agentic coding, computer use, and long-horizon knowledge work. On the ARC-AGI-3 novel problem-solving benchmark, Opus 5 scores 30.2%, roughly four times GPT-5.6 Sol's 7.8% and twenty times Opus 4.8's 1.5%. On the GDPval-AA v2 knowledge-work benchmark, it reaches an Elo of 1,861, ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736). Priced at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8, it offers a 1,000,000-token context window, 128,000-token max output, and tool calling. It fits developers building coding agents and automation pipelines that need strong reasoning at Opus-level pricing.

by Anthropic | $5/M input | $25/M output

Claude Opus 5 Fast is a high-speed configuration of Anthropic's Opus 5 flagship model, generating output tokens at roughly 2.5x the speed of standard Opus 5. It runs the same Opus 5 model, which Anthropic says outperforms Opus 4.8 on agentic coding, computer use, and knowledge work, and approaches Fable 5's performance at roughly half the cost. Fast mode pricing is $10/$50 per million input/output tokens, twice the price of standard Opus 5 ($5/$25) and in line with what Opus 4.8 Fast cost. It supports the full 1M token context window and 128k max output tokens. Choose Opus 5 Fast for latency-sensitive agentic pipelines, live coding sessions, and real-time workflows where throughput matters. For cost-sensitive or batch workloads, standard Opus 5 offers the same intelligence at half the price.

by Anthropic | $10/M input | $50/M output

Ling 3.0 Flash is InclusionAI's (Ant Group) successor to Ling 2.6 Flash, a hybrid-reasoning Mixture-of-Experts model with 124B total parameters and about 5.1B active per token. It stacks five Kimi Delta Attention (KDA) layers per one Multi-Head Latent Attention (MLA) layer, combining efficient long-range memory with precise attention, and supports both thinking and non-thinking modes. According to InclusionAI, with roughly 1/8 of the total parameters and 1/12 of the active parameters of its 1T-parameter flagship model, Ling 3.0 Flash matches or beats that flagship on most of the benchmarks the company reported. It natively supports a 262K-token context window, with InclusionAI designing it to scale toward 1M context; the paid tier here is served at 131K context. It targets high-frequency agentic workflows such as coding agents, tool use, document processing, and long multi-turn conversations, where per-token cost and latency matter.

by InclusionAI | $0.06/M input | $0.18/M output

Laguna S 2.1 is Poolside's mid-size agentic coding model, sitting between the Laguna XS and M families. It's a 118B-parameter Mixture-of-Experts model with 8B activated parameters, released under the OpenMDW-1.1 license with a 1M-token context window. It scores 78.5% on SWE-bench Multilingual, 59.4% on SWE-bench Pro, and 70.2% on Terminal-Bench 2.1. Poolside reports these results match or exceed models with two to eight times as many active parameters, including DeepSeek-V4-Flash and NVIDIA's Nemotron 3 Ultra. Laguna S 2.1 supports tool calling, thinking mode, and multi-step reasoning for agentic coding workflows. It suits developers who need stronger coding performance than XS 2.1 without the cost of the larger M.1 model.

by Poolside | $0.09/M input | $0.18/M output

Gemini 3.5 Flash-Lite is Google's fastest and most cost-efficient model in the Gemini 3.5 series, built for high-throughput, low-latency workloads. It scores 54% on Terminal-Bench 2.1 and 72.2% on GDM-MRCR v2, up from 31% and 60.1% for Gemini 3.1 Flash-Lite. It also outperforms the larger Gemini 3 Flash on SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%), while generating output at roughly 350 tokens per second. It supports text, image, video, audio, and PDF input with a 1M token context window, configurable thinking levels, and function calling, including computer use as a built-in tool. It's suited for agentic search, document processing, and other high-volume tasks where throughput and cost matter more than maximum reasoning depth.

by Google | $0.15/M input | $1.25/M output

Gemini 3.6 Flash is Google's workhorse Flash-tier model, released as the successor to Gemini 3.5 Flash. It's built for running AI agents in production, with improvements in coding precision, computer use, and multimodal understanding. In Google's own benchmarks, it scores 83.0% on OSWorld-Verified (up from 78.4% for 3.5 Flash), 49% on DeepSWE (up from 37%), 63.9% on MLE-Bench (up from 49.7%), and 58.7% on SWE-Bench Pro. It also produces 17% fewer output tokens than 3.5 Flash on comparable tasks. It accepts text, image, video, audio, and PDF input with a 1M token context window, supports function calling and a built-in computer-use tool, and has a March 2026 knowledge cutoff. At $0.75 per million input tokens and $3.75 per million output tokens, it's well under 3.5 Flash's $1.50 and $9.00 rates.

by Google | $0.75/M input | $3.75/M output

LongCat 2.0 is a 1.6-trillion-parameter Mixture-of-Experts (MoE) language model from Meituan, activating between roughly 33B and 56B parameters per token depending on query complexity. It supports a native 1-million-token context window through a custom LongCat Sparse Attention mechanism built for long-context workloads, and is tuned for agentic coding tasks such as whole-repository reasoning, multi-step terminal use, and cross-language migration. On SWE-bench Pro it scores 59.5, edging GPT-5.5's 58.6, and reaches 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-bench Multilingual; Meituan reports overall performance comparable to Google's Gemini 3.1 Pro. A good fit for developers building coding agents or tools that need large context windows for whole-codebase reasoning.

by Meituan | $0.75/M input | $2.95/M output

Kimi K3 is Moonshot AI's flagship open-weight model, released July 16, 2026, with full weights following on July 27. At roughly 2.8 trillion parameters in a Mixture-of-Experts architecture, Moonshot positions it as the largest open-source model released to date, built on two new components: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, a replacement for standard residual connections. It runs in an always-on thinking mode with a 1-million-token context window and accepts text, image, and video input. Reported results include 93.5% on GPQA Diamond, 91.2% on BrowseComp, 88.3% on Terminal-Bench 2.1, and a first-place finish on Arena.ai's Frontend Code Arena, putting it close to Claude Opus 4.8 and GPT-5.5 on several agentic and coding tasks. These figures come from Moonshot and early testers, not independently confirmed leaderboards. It suits developers building long-horizon coding agents and tool-calling pipelines who want frontier-level performance at open-weight pricing.

by Moonshot AI | $3/M input | $15/M output

Muse Spark 1.1 is a multimodal reasoning model from Meta Superintelligence Labs, built for agentic workflows. It accepts text, images, video, audio, and PDF documents as input and returns text, with a 1,048,576-token context window. The model is designed to orchestrate multi-agent workflows, acting as either a main agent that plans and delegates tasks or as a subagent, and it generalizes zero-shot to new tools, MCP servers, and custom skills. It supports parallel function calling, structured output, built-in search with citations, and configurable reasoning effort, and Meta reports strong results on coding across large codebases, computer-use tasks, and visual-to-code generation. This is Meta's first model available through a paid API, priced at $1.25 per million input tokens and $4.25 per million output tokens, aimed at developers building agentic coding tools and enterprise workflow automation.

by Meta Llama | $1.25/M input | $4.25/M output

KAT-Coder-Pro V2.5 is Kwaipilot's (Kuaishou's AI research division) successor to KAT-Coder-Pro V2, an agentic coding model built on a Mixture-of-Experts architecture with 72B active parameters. It offers a 256K token context window with up to 80K output tokens, and is trained through large-scale agentic reinforcement learning using reconstructed, verifiable repository environments rather than scale alone. On repository-level software engineering benchmarks, KAT-Coder-Pro V2.5 scores 65.2 on SWE-Bench Pro, ranking second only to Opus 4.8 (69.2). It posts the best agentic tool-use result among tested models on PinchBench (94.9), reflecting strength in long-horizon, multi-step tasks. The model is designed to take on an entire issue or business workflow autonomously, locating problems, making changes, and completing tasks inside real repositories. It also has a lighter sibling variant, KAT-Coder-Air V2.5, for lower-cost workloads.

by KwaiPilot | $0.74/M input | $2.96/M output

KAT-Coder-Air V2.5 is a coding-focused agentic model from Kwaipilot (Kuaishou's AI research division), released July 10, 2026 as the faster, lower-cost sibling to KAT-Coder-Pro V2.5. It shares Pro's 256K context window and 80K max output tokens, priced at $0.15 per 1M input tokens and $0.60 per 1M output tokens, roughly a fifth of Pro's $0.74/$2.96 rates. The model is built for agent and tool-use workflows inside real, executable repositories, handling issue localization, code modification, and test execution as part of an end-to-end development loop. Kwaipilot designed it for interactive development and rapid iteration rather than single-shot generation, with integration into toolchains such as Claude Code and OpenHands. It supports function calling, tool use, structured JSON output, and prompt caching. KAT-Coder-Air V2.5 targets teams running high-volume coding agents and workflow automation where speed and cost per request matter, while KAT-Coder-Pro V2.5 remains Kwaipilot's higher-capability option for complex software engineering tasks.

by KwaiPilot | $0.15/M input | $0.6/M output

GPT-5.6 Sol is OpenAI's flagship model in the GPT-5.6 family, sitting above Terra and Luna as the most capable and most expensive of the three. It reached general availability on July 9, 2026. Sol is built for agentic coding, cybersecurity research, and long-horizon autonomous work. OpenAI reports it sets a new state of the art on Terminal-Bench 2.1, a benchmark for command-line workflows that require planning, iteration, and tool coordination, and describes it as the most capable model yet for cybersecurity tasks such as vulnerability research. The GPT-5.6 family adds a max reasoning effort setting, an ultra mode that coordinates subagents on complex tasks, and Programmatic Tool Calling in the Responses API. Sol carries a 1.05M token context window, 128,000 max output tokens, and a February 16, 2026 knowledge cutoff, making it best suited for teams that need the highest available capability and can absorb its premium per-token cost.

by OpenAI | $2.5/M input | $15/M output

GPT-5.6 Sol Pro is OpenAI's highest-capability configuration of GPT-5.6 Sol, the flagship tier in the GPT-5.6 family alongside the smaller Terra and Luna models. Sol is built for complex reasoning, coding, scientific work, and long-running agentic tasks. Sol Pro is not a separate, larger model. It runs the same underlying model as base Sol with reasoning mode set to pro for higher-quality responses on harder problems, which is why it shares identical pricing with the base model, $5 per million input tokens and $30 per million output tokens, unlike GPT-5.4 Pro and GPT-5.5 Pro, which cost several times more than their base models. It carries the same 1,050,000-token context window and 128,000 max output tokens as GPT-5.6 Sol, and supports image input, function calling, and the Responses API tool suite. It launched July 9, 2026, aimed at the hardest, longest-running tasks where response quality matters more than cost or latency.

by OpenAI | $5/M input | $30/M output

GPT-5.6 Terra is OpenAI's mid-tier chat model in the GPT-5.6 family, positioned between the flagship Sol and the faster, cheaper Luna. OpenAI's documentation describes it as designed for workloads that balance intelligence and cost, corresponding to the mini tier used in earlier GPT-5 families. In OpenAI's naming system, the number marks a model's generation, while Sol, Terra, and Luna mark capability tiers that can each advance on their own schedule. It has a 1,050,000 token context window, up to 128,000 output tokens, and a February 16, 2026 knowledge cutoff. Pricing is $2.50 per million input tokens and $15 per million output tokens, between Luna's $1/$6 and Sol's $5/$30. Terra accepts text and image input, supports function calling and tool use, and is aimed at high-volume business tasks such as customer support, internal tools, and document analysis, alongside everyday interactive and agentic coding where Sol's higher reasoning ceiling isn't needed.

by OpenAI | $1/M input | $6/M output

GPT-5.6 Terra Pro is OpenAI's higher-effort reasoning configuration of GPT-5.6 Terra, the mid-tier model in the GPT-5.6 family that sits between the flagship Sol and the economical Luna. It is the same underlying model as base Terra, run with reasoning mode set to pro for higher-quality responses on complex tasks, rather than a separate, larger checkpoint. Terra Pro is priced at $1 per million input tokens and $6 per million output tokens, compared to base Terra's $2.50/$15 pricing. OpenAI positions Terra as offering performance competitive with GPT-5.5 at a lower cost, aimed at high-volume business tasks such as customer support, internal tools, and document analysis. It supports a 1.05M token context window, image input, and function calling, making Terra Pro a fit for teams that want more careful reasoning on harder requests within that workload without paying Sol-level prices.

by OpenAI | $1/M input | $6/M output

GPT-5.6 Luna Pro is OpenAI's extended-reasoning configuration of GPT-5.6 Luna, the fastest and lowest-cost tier in the three-model GPT-5.6 family alongside Terra and Sol, released July 2026. Rather than a separate, larger model, Luna Pro is the same underlying Luna model run with its reasoning mode set to pro, trading speed for more reasoning time on harder requests. It is priced at $0.10 per million input tokens and $0.60 per million output tokens, below base Luna's $1/$6 pricing. It carries a 1,050,000 token context window, 128,000 max output tokens, and support for function calling. Best suited for teams already using Luna for high-volume, latency-sensitive work such as summarization, drafting, and routine automation who need higher accuracy on a subset of harder requests without switching models or providers.

by OpenAI | $0.1/M input | $0.6/M output

GPT-5.6 Luna is OpenAI's smallest and most cost-efficient model in the GPT-5.6 family, positioned below GPT-5.6 Terra and the flagship GPT-5.6 Sol. It targets high-volume, cost-sensitive workloads such as customer service automation, content moderation, and data extraction, where lower per-token cost matters more than peak capability. It shares the same 1,050,000-token context window, 128,000 max output tokens, and February 16, 2026 knowledge cutoff as its larger siblings. Luna accepts text and image input, produces text output, and supports function calling, structured outputs, streaming, and reasoning tokens, along with tool access including web search, file search, code interpreter, computer use, and MCP. At $1 per million input tokens and $6 per million output tokens, Luna costs 60% less on input than GPT-5.6 Terra, making it suited for applications that run at scale rather than tasks requiring the highest accuracy.

by OpenAI | $0.1/M input | $0.6/M output

Grok 4.5 is xAI's flagship mixture-of-experts model, trained jointly with Cursor on trillions of tokens of real coding sessions alongside STEM, research, and knowledge-work data. Elon Musk described it as an Opus-class model, faster and more token-efficient at a lower cost than comparable frontier models. It is tuned for coding, including Rust and C/C++, and for multi-step agentic workflows. It also handles finance and legal work, and can build multi-sheet Excel models with live web research, design diagrams in PowerPoint, and write structured documents in Word. Reasoning effort is configurable as low, medium, or high, with high as the default. The model accepts text and image input, supports a 500,000 token context window, and includes native tool calling. At $2 per million input tokens and $6 per million output tokens, it suits developers building coding agents and knowledge-work tools who need frontier-level capability without paying Opus-class pricing.

by xAI | $2/M input | $6/M output

Aion-3.0 is a multi-model roleplaying and storytelling system from AionLabs, built on the GLM family of models. Rather than a single model generating the full response, it runs a collaborative generation process in which multiple specialized models each contribute, aimed at producing stronger narrative structure and more compelling tension and conflict. It handles mature and darker themes with notable nuance, continuing the roleplay and narrative focus of AionLabs' earlier Aion models such as Aion-2.0 and Aion-RP. The model supports a 131K-token context window with up to 32K tokens of output. Aion-3.0 is a good fit for developers building interactive fiction, character-driven chat, or other creative-writing tools where narrative depth and thematic nuance matter more than raw speed.

by Aion Labs | $3/M input | $6/M output

Aion-3.0-Mini is a multi-model roleplaying and storytelling system from AionLabs, built on the DeepSeek family of models rather than the GLM base used in the full Aion-3.0. Like its larger sibling, it runs a collaborative generation process in which multiple specialized models each contribute to a response, aimed at producing stronger narrative structure and more compelling tension and conflict. It handles mature and darker themes with notable nuance, continuing the roleplay and narrative focus of AionLabs' earlier Aion models. The model supports a 131K-token context window with up to 32K tokens of output. Pricing is $0.70 per million input tokens and $1.40 per million output tokens, roughly a quarter of Aion-3.0's cost, making it a fit for developers who want the same collaborative-generation approach to character-driven chat and interactive fiction at a lower price point.

by Aion Labs | $0.7/M input | $1.4/M output

Tencent Hy3 is the official release of Tencent Hunyuan's Hy3 series, a hybrid fast-and-slow-thinking Mixture-of-Experts model with 295B total parameters and 21B active per query, following the earlier hy3-preview. It supports a 262K-token context window and up to 131K output tokens. Compared to the preview, Tencent reports agent and coding capability gains of 20-30%, a hallucination rate cut from 12.5% to 5.4%, and commonsense error rates nearly halved. It scores 78 on SWE-bench Verified, 57.9 on SWE-bench Pro, and 90.4 on GPQA Diamond, and Tencent says it matches flagship models with two to five times its parameter count. Hy3 ships under Apache 2.0 and already powers Tencent products like WorkBuddy/CodeBuddy and Yuanbao, with a reported 90% task resolution rate on Tencent's internal WorkBuddy platform. It's a good fit for developers building coding agents, tool-using workflows, and long-context reasoning pipelines who want strong performance at a lower active-parameter cost.

by Tencent | $0.14/M input | $0.58/M output

Laguna XS 2.1 is Poolside's updated compact agentic coding model, a 33B-parameter Mixture-of-Experts architecture with 3B activated parameters, released under the OpenMDW-1.1 license as a successor to Laguna XS.2. It scores 70.9% on SWE-bench Verified, 63.1% on SWE-bench Multilingual, 47.6% on SWE-bench Pro, and 37.5% on Terminal-Bench 2.0, each an improvement over XS.2, with the largest gain (+5.4 points) on the multilingual benchmark. It trails larger models like Qwen3.6-35B-A3B on multilingual coding but stays competitive within its size class. The model keeps the same 262K context window, tool calling, and reasoning support as its predecessor. It suits developers already using Laguna XS.2 who want incremental gains on multilingual and terminal-style coding tasks without moving to a larger model.

by Poolside | $0.06/M input | $0.12/M output

Gemini 3.1 Flash Lite Image is Google DeepMind's fastest and lowest-cost image generation and editing model in the Gemini 3.1 family, also known as Nano Banana 2 Lite. It generates a 1K-resolution image in about 4 seconds, roughly 2.7x faster than Gemini 3.1 Flash Image, at $0.0336 per image. It keeps the character consistency, precise natural-language editing, and real-world knowledge of the Nano Banana family while trading some peak quality for speed. It supports text-to-image generation, image editing, and multi-image composition through the same API used by other Gemini image models, with a discrete set of 14 aspect ratios at 1K resolution. It's built for high-volume, latency-sensitive use cases such as rapid prototyping, interactive applications, and batch image generation, where cost and turnaround time matter more than maximum fidelity.

by Google | $0.0336/image

Claude Sonnet 5 is Anthropic's newest Sonnet-tier model, released June 30, 2026 as the successor to Sonnet 4.6. Anthropic built it to be the most agentic Sonnet yet, with gains in planning, tool use, coding, and multi-step reasoning that bring it close to Opus 4.8 quality at a lower price. On SWE-bench Pro it scores 63.2%, up from Sonnet 4.6's 58.1% and ahead of GPT-5.5's 58.6%, though still behind Opus 4.8's 69.2%. On Terminal-Bench 2.1 it reaches 80.4% versus Sonnet 4.6's 67.0%, and on the OSWorld-Verified computer-use benchmark it posts 81.2% against Sonnet 4.6's 78.5%. It targets developers running agentic coding pipelines, automation workflows, and tool-heavy tasks who want most of Opus-tier capability without Opus pricing. Introductory pricing is $2/$10 per million input/output tokens through August 31, 2026, after which it reverts to the standard $3/$15 listed here.

by Anthropic | $3/M input | $15/M output

Dola Seed 2.1 Turbo is ByteDance's faster, lower-cost variant in the Seed 2.1 family, offered internationally through BytePlus's Model Ark platform. It accepts text, image, and video input, with a 256K context window and function calling for agent workflows. On ByteDance's own benchmarks against Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro, Turbo scores 89 on VideoMME for long video understanding and 79.2 on OVOBench for streaming video, ahead of Gemini 3.1 Pro's 86.7 and 64.1. It scores 88.0 on BeyondAIME reasoning, above Seed 2.1 Pro's 87.0. On knowledge tests like SuperGPQA it trails GPT-5.5 and Gemini. These are vendor-reported figures, not independently verified. It suits high-volume API use such as video analysis, document processing, and agent pipelines, where cost and video handling matter more than peak coding scores.

by BytePlus | $0.5/M input | $2.5/M output

Dola Seedream 5.0 Pro is ByteDance's Seedream 5.0 Pro image generation and editing model, offered here through BytePlus's ModelArk API under the "Dola" branding BytePlus uses across its hosted model lineup. It generates images as independently editable layers, separating subjects, text, and background so pieces can be repositioned or swapped after generation. It supports pixel-level editing through selection points, arrows, and annotation boxes, reads color codes and material specifications, and renders text across 14 or more languages including Arabic, Korean, Thai, French, Russian, and Japanese. On Artificial Analysis's independent Text-to-Image Elo leaderboard it scores 1,283, ranking 8th overall, ahead of Black Forest Labs' FLUX.2 Max (1,233) and just behind Google's Nano Banana Pro (1,299). It fits production teams that need layered, editable output and multilingual text rather than a single flattened image.

by BytePlus | $0.045/image

Grok Imagine Image is xAI's standard text-to-image generation model, built on Aurora — an autoregressive Mixture-of-Experts architecture trained on billions of text-image pairs. It accepts text prompts and optional reference images as input, producing up to 10 images per request at 1K (1024×1024) or 2K (2048×2048) resolution across 13 aspect ratios. Output formats include JPEG, PNG, and WebP. The model is noted for strong instruction following, handling style transfer, object addition or removal, and multi-reference composition through natural language alone. It generates images quickly, making it practical for high-volume pipelines. Best suited for product mockups, marketing visuals, social media graphics, and concept art prototyping where speed and prompt adherence matter. A higher-quality variant, Grok Imagine Image Pro, is available when output fidelity is the priority.

by xAI | $0.02/image

Grok Imagine Image (Quality) is xAI's high-fidelity text-to-image generation model, built on the Aurora autoregressive Mixture-of-Experts architecture rather than a diffusion backbone. It targets applications where output quality is the priority over speed — delivering more natural lighting, richer textures, and cinematic consistency compared to standard-tier image models. Text rendering in generated images is notably legible across multiple languages, addressing a historically weak area for generative image models. The API supports 1K and 2K resolutions, 13 aspect ratios, and JPEG, PNG, and WebP outputs. Up to four image candidates can be generated per request. At $0.05 per image, it is well suited for product visuals, marketing assets, brand imagery, and creative workflows that require production-ready quality.

by xAI | $0.05/image

Fugu Ultra is a learned multi-agent orchestration system from Tokyo-based Sakana AI that routes tasks across a swappable pool of frontier LLMs behind a single OpenAI-compatible endpoint. Rather than relying on one monolithic model, Fugu Ultra dynamically assigns Thinker, Worker, and Verifier roles to specialist agents, then synthesizes their outputs into a single response. The underlying coordination is grounded in Sakana's TRINITY and Conductor research, published at ICLR 2026. It targets demanding, multi-step problems: complex reasoning, code review, agentic workflows, cybersecurity analysis, and research tasks. On LiveCodeBench it scores 93.2, ahead of several frontier competitors, and it matches leading models on GPQA-Diamond and Humanity's Last Exam. Fugu Ultra is a strong choice for developers who need frontier-level quality on hard tasks without committing to a single model provider.

by Sakana AI | $5/M input | $30/M output

Nex N2 Mini is Nex AGI's smaller agentic model in the Nex N2 family, built on a Qwen3.5 base with 35B total parameters and 3B active per token, released open-source under Apache 2.0 alongside the larger Nex N2 Pro. It shares Nex N2 Pro's Agentic Thinking framework, pairing Adaptive Thinking, deciding on its own when and how deeply to reason, with Coherent Thinking, a consistent reasoning approach carried across general and agentic tasks. It supports a 262K context window and function calling. On benchmarks it scores 50.2 on SWE-Bench Pro, 60.7 on Terminal-Bench 2.1, 82.6 on GPQA Diamond, and 74.1 on BrowseComp, below Nex N2 Pro's 58.8, 75.3, and 90.7 on the same tests but at a fraction of the price. It fits developers who want Nex N2's agentic coding and tool-calling behavior for lighter, latency-sensitive, or budget-constrained workloads rather than maximum accuracy.

by Nex AGI | $0.03/M input | $0.1/M output

GLM 5.2 Fast is the high-speed serving tier of Z.ai's GLM-5.2, running the same model weights on inference infrastructure tuned for higher throughput. Output quality matches the standard GLM-5.2 endpoint; serving speed and price are the differences. It keeps the full 1M-token context window and 128K max output, along with tool calling, structured output, streaming, optional thinking mode with adjustable reasoning effort, and implicit prompt caching. Providers report roughly 2x the throughput of their standard GLM-5.2 endpoints, with peaks measured at 446 tokens per second. Because the weights are identical, it inherits GLM-5.2's coding results, including 62.1 on SWE-bench Pro. The tradeoff is a higher per-token price than standard GLM-5.2. A fit for agent loops that chain many model calls, real-time coding assistants, and other latency-sensitive workloads.

by Z.AI | $2.29/M input | $8/M output

Seed 2.1 Turbo is the faster, lower-cost variant in ByteDance's Seed 2.1 family, a model line built for coding, long-horizon agent tasks, and multimodal understanding. It accepts text and image input, has a 256K context window, and supports function calling in an OpenAI-compatible format, so existing tool-use pipelines port over with little change. ByteDance positions it for high-volume production workloads such as agent pipelines, coding assistance, and document processing, and prices it at roughly half the cost of Seed 2.1 Pro. The company reports performance close to the Pro version and comparable to GPT-5.5 on coding and agent tasks, though no independently verified benchmark scores for Turbo were published at launch. Choose it when volume and cost drive the decision rather than peak capability; Seed 2.1 Pro remains the stronger option for the hardest tasks.

by ByteDance | $0.5/M input | $2.5/M output

Cohere North Mini Code is an open-weight agentic coding model from Cohere Labs, built on a sparse Mixture-of-Experts architecture with 30B total parameters and 3B active per token. It is optimized for code generation, agentic software engineering workflows — including sub-agent orchestration, multi-file analysis, and code review — and real terminal tasks. On the Artificial Analysis Coding Index it scores 33.4, outperforming open-source models up to 4x its size. It delivers 2.8x higher throughput and 30% lower inter-token latency than Devstral Small 2. With a 256K context window and 64K max output, it suits developers building AI coding agents or automated engineering pipelines. Released under Apache 2.0.

by Cohere | $0/M input | $0/M output

GLM-5.2 is an open-weight Mixture-of-Experts model from Zhipu AI (Z.ai), available here through BytePlus's hosted API. The model has roughly 744 billion total parameters and a 1 million token context window, and Z.ai released its weights under an MIT license. GLM-5.2 is built for coding, particularly long-horizon and agentic coding tasks. On SWE-bench Pro it scored 62.1, ahead of GPT-5.5 (58.6) and its predecessor GLM-5.1 (58.4). On the Artificial Analysis Intelligence Index it scored 51, the highest score among open-weight models at release. It suits teams building coding agents or working with large codebases who want an open-weight model, accessed here through BytePlus's API instead of self-hosting.

by BytePlus | $1.4/M input | $4.4/M output

GLM 5.2 is Z.ai's sixth-generation open-weight foundation model, built around a 1M-token context window that can hold entire mid-sized codebases in a single prompt. It uses a 744B-parameter Mixture-of-Experts architecture with an "IndexShare" attention optimization that cuts per-token FLOPs by 2.9x at 1M context, keeping long-context inference practical. A new MTP speculative decoding layer raises acceptance length by up to 20%, reducing latency. Dual reasoning modes (High/Max) let you trade speed for depth on complex tasks. GLM 5.2 scored 62.1 on SWE-bench Pro, outperforming GPT-5.5 (58.6) and its predecessor GLM-5.1 (58.4). It is the top-ranked open-weight model on long-horizon coding benchmarks. Best suited for repository-scale refactoring, multi-step agentic coding, full-codebase analysis, and any workflow that previously required chunking large inputs.

by Z.AI | $0.55/M input | $1.93/M output

Kimi K2.7 Code Fast is the high-speed serving tier of Kimi K2.7 Code, Moonshot AI's coding-agent model. The weights and capabilities are the same as the standard route; the difference is throughput. Moonshot reports around 180 tokens per second on coding tasks with median-length inputs and up to 260 tokens per second on shorter-context tasks, up to 6x faster than the standard release. Like the base model, it runs in always-on thinking mode with a 262K-token context window and supports tool calling for agentic workflows. Compared with K2.6, K2.7 Code improves coding and agent performance while using roughly 30% fewer thinking tokens. The fast route costs about twice as much per token as the standard endpoint, so it fits interactive coding agents and IDE assistants where response speed matters more than cost.

by Moonshot AI | $1.86/M input | $7.72/M output

Dreamina Seedance 2.0 Mini is the smallest, most economical tier of ByteDance's Seedance 2.0 video family, distributed under the Dreamina brand and hosted here via BytePlus. BytePlus positions it as roughly twice as fast as the Fast tier and about half the list price of the full Seedance 2.0 model, while still accepting image, video, and audio references alongside a text prompt. Output is capped at 720p and 480p with clips up to 15 seconds at 24fps, the same resolution ceiling as the Fast tier. ByteDance markets it for everyday production work such as brainstorming, quick mockups, and short-form video, where cost and turnaround matter more than peak quality.

by BytePlus

Kimi K2.7 Code is Moonshot AI's open-weight coding-agent model, released June 2026 and purpose-built for long-horizon, autonomous coding tasks. It shares the same 1-trillion-parameter Mixture-of-Experts architecture (32B active parameters) as K2.6 but is entirely focused on software engineering workloads. Compared to K2.6, it improves 21.8% on Kimi Code Bench v2, 11% on Program Bench, and 31.5% on MLS Bench Lite, while cutting reasoning-token usage by roughly 30%. It always runs in thinking mode — non-thinking mode is not supported. With a 262K-token context window, K2.7 Code is well-suited for multi-file, repository-scale coding pipelines and agentic workflows where sustained reasoning and deep code understanding matter.

by Moonshot AI | $0.45/M input | $1.86/M output

Claude Fable 5 is Anthropic's most capable generally available model — a publicly released, safety-hardened version of its frontier Mythos-class system, launched June 2026. It is state-of-the-art on nearly every benchmark Anthropic tested, with its lead over prior Claude models growing the longer and more complex the task. It scores 80.3% on SWE-bench Pro, well ahead of Claude Opus 4.8 (69.2%) and GPT-5.5 (58.6%). During early testing, Stripe used it to run a codebase-wide migration in a day that would have taken a team over two months by hand. It's also Anthropic's new state-of-the-art for vision tasks and can work autonomously longer than any previous Claude model. Priced at $10/$50 per million input/output tokens — less than half the cost of Claude Mythos Preview — it targets developers building agentic coding pipelines, vision-heavy workloads, and long-horizon autonomous tasks. Queries on a small set of sensitive topics fall back to Opus 4.8 (on average under 5% of sessions).

by Anthropic | $10/M input | $50/M output

Nex N2 Pro is an agentic mixture-of-experts model from Nex AGI, built on a Qwen3.5 base with 397B total parameters and 17B active per token, released open-source under Apache 2.0 in June 2026. The model's Agentic Thinking framework closes the loop between planning, code implementation, environmental feedback, and iterative debugging — making it especially capable at long-horizon tasks. It supports function calling, structured outputs, vision inputs, and reasoning traces. Benchmark highlights include 75.3 on Terminal-Bench 2.1, 80.8 on SWE-Bench Verified, and 90.7 on GPQA Diamond. It is a strong fit for developers building coding agents, deep research pipelines, or multi-step automation that demands reliable tool use and a 262K context window.

by Nex AGI | $0.25/M input | $1/M output

Nemotron 3 Ultra 550B A55B is NVIDIA's open-weight frontier reasoning model with 550B total and 55B active parameters, built on a hybrid Mamba-Transformer Mixture-of-Experts architecture. It supports a 1M token context window and is designed for long-running agentic workflows, complex multi-step reasoning, and high-accuracy tasks across code, math, and science. NVIDIA reports up to 5.9x higher inference throughput than comparable open MoE models. On the Artificial Analysis Intelligence Index it scores 48, leading US open-weight models and delivering the highest non-hallucination score in its comparison set (78.7 on AA-Omniscience). Choose it for production agentic pipelines, deep document analysis, or reasoning-heavy API workloads where both accuracy and throughput matter.

by NVIDIA | $0.5/M input | $2.5/M output

Nemotron 3.5 Content Safety is a compact 4B-parameter multimodal guardrail model from NVIDIA, fine-tuned from Google Gemma-3-4B, designed to moderate both inputs and outputs of LLMs and VLMs. It classifies prompts and responses as safe or unsafe across 23 safety categories based on the Aegis v2 taxonomy, supports 12 languages, and accepts both text and image input. An optional reasoning mode provides step-by-step chain-of-thought traces explaining each decision — useful for auditing and policy tuning. Despite its 4B size, it leads external multimodal safety benchmarks including the top harmful-F1 score on VLGuard, matching or beating 8–12B models. It also supports custom operator-defined content policies enforced at inference time. Choose it for prompt and response moderation pipelines, safety evaluation of LLM outputs, or as an inference-time guardrail in enterprise AI applications requiring explainable, policy-aware content filtering.

by NVIDIA | $0/M input | $0/M output

Qwen3.7 Plus is Alibaba's multimodal agent model, released in June 2026, combining vision-language understanding with full agentic capabilities across a 1 million-token context window. Unlike the text-only Qwen3.7 Max, Plus ingests images and video alongside text, processed through early-fusion training so vision and language are jointly understood from the first layer. This enables GUI grounding — the model can interpret screenshots and issue precise on-screen actions — scoring 79.0 on ScreenSpot Pro, placing it alongside Claude Computer Use and OpenAI Operator in the GUI automation tier. Beyond vision, it adds deep reasoning, self-programming, tool invocation, and autonomous iteration: the model writes and tests code, calls external APIs, and loops until the task is done. On the Artificial Analysis Intelligence Index it scores 53. Choose it over Qwen3.7 Max when your workflow requires image or video inputs, browser/desktop automation, or end-to-end agentic pipelines that combine seeing, reasoning, and doing.

by Qwen | $0.28/M input | $1.1/M output

Ideogram 4.0 is an open-weight text-to-image model from Ideogram AI, released June 3, 2026 as a 9.3B-parameter diffusion transformer with weights published under Apache 2.0. It supports structured JSON prompting with bounding-box layout and hex color control, native 2K output, and an Ideogram Character feature for subject consistency from one reference photo. It continues Ideogram's focus on text rendering, scoring 0.97 OCR accuracy on X-Omni and 0.69 mIoU on 7Bench for layout control. In Ideogram's designer-preference evaluation (4,366 votes across 9 models), it scored an ELO of 1062, second overall behind GPT Image 2 and first among open-weight models, ahead of larger models like Qwen-Image (20B) and HunyuanImage 3.0 (80B MoE). API pricing runs $0.03-$0.10 per image by quality tier, with Style Reference (up to three images), Remix, Edit, and Reframe support. It suits developers building advertising, branding, packaging, and product-photography pipelines needing typography and layout control.

by Ideogram | $0.06/image

MiniMax M3 is a frontier-level multimodal language model built for long-horizon coding, agentic workflows, and complex reasoning. It introduces MiniMax Sparse Attention (MSA), which delivers 15x faster decoding and 9x faster prefill at 1M-token context compared to the prior generation. M3 scores 59.0% on SWE-Bench Pro — surpassing GPT-5.5 and Gemini 3.1 Pro — and achieves the highest score on Claw-Eval (74.5%) for autonomous agent tasks. It accepts text, image, and video inputs natively and supports tool calling through standard MCP scaffolding. Ideal for developers building coding agents, long-document pipelines, and multi-step automation that require sustained, multi-hour autonomous execution.

by MiniMax | $0.28/M input | $1.1/M output

Step 3.7 Flash is a multimodal Mixture-of-Experts model from StepFun, combining a 196B-parameter language backbone with a 1.8B vision encoder and activating roughly 11B parameters per token. It natively processes text and images, supporting a 256K-token context window. The model is designed for agentic coding, tool orchestration, long-context reasoning, and search-intensive workflows. It exposes tunable reasoning tiers (low, medium, high) so developers can trade inference cost against answer depth on a per-call basis. Compared to Step 3.5 Flash, Step 3.7 Flash adds native vision input, improves cross-harness consistency, and achieves stronger performance on coding agent and terminal benchmarks. It targets developers building production pipelines that combine perception, retrieval, and multi-step execution.

by StepFun | $0.2/M input | $1.15/M output

Claude Opus 4.8 is Anthropic's latest flagship AI model, building on Opus 4.7 with improvements across coding, reasoning, and knowledge work — at the same price point. The model excels at complex, multi-step agentic workflows including coding, financial analysis, legal reasoning, and browser-based computer use. Early testers report it outperforms both prior Opus models and GPT-5.5 across several agentic benchmarks. Notably, it's around four times less likely than Opus 4.7 to let flaws in generated code pass without flagging them. Opus 4.8 supports configurable effort levels, letting developers trade quality against speed and token cost. A fast mode operates at 2.5× speed. Priced at $5/$25 per million input/output tokens, it targets teams running autonomous agents, analysis pipelines, and long-running async workflows where reliability and judgment are critical.

by Anthropic | $5/M input | $25/M output

Claude Opus 4.8 Fast is a high-speed configuration of Anthropic's latest flagship model, delivering up to 2.5x faster output token generation at significantly lower cost than previous fast variants. It runs the same Opus 4.8 model — which improves on Opus 4.7 in agentic coding (69.2% vs 64.3%), multidisciplinary reasoning with tools (57.9% vs 54.7%), and knowledge work — but optimized for lower latency. Fast mode pricing is $10/$50 per million input/output tokens, three times cheaper than fast mode for Opus 4.6 and 4.7 ($30/$150). It supports the full 1M token context window and 128k max output tokens. Choose Opus 4.8 Fast for latency-sensitive agentic pipelines, live coding sessions, and real-time workflows where throughput matters. For cost-sensitive or batch workloads, standard Opus 4.8 offers the same intelligence at half the price.

by Anthropic | $10/M input | $50/M output

Qwen3.7 Max is Alibaba's flagship proprietary reasoning model, released in May 2026, built for long-horizon agentic workloads with a 1 million-token context window and a chain-of-thought reasoning architecture. It is purpose-built for complex, multi-step autonomous tasks. Alibaba demonstrated the model running for 35 hours without degradation, executing over 1,000 tool calls in a single session — making it a strong candidate for coding agents, automated pipelines, and deep document analysis. On benchmarks, it ranks 13th globally on LM Arena's text leaderboard and scores 56.6 on the Artificial Analysis Intelligence Index, making it the highest-ranked Chinese model on that index. It posted 90.2 on Arena-Hard v2 and 72.5 on SWE-Bench Verified. Qwen3.7 Max supports the Anthropic API protocol natively, so it integrates cleanly with tooling like Claude Code. It is well-suited for developers building coding assistants, research agents, or any API use case requiring extended reasoning over large contexts.

by Qwen | $1.65/M input | $4.95/M output

Grok Build 0.1 is xAI's fast coding model trained specifically for agentic software engineering workflows. Released in May 2026 and currently in early access, it is purpose-built for interactive coding agents, tool use, and multi-step development tasks rather than general-purpose conversation. The model accepts text and image inputs and produces text output, with a 256,000-token context window. It supports function calling, structured outputs, and built-in reasoning that is always active, enabling it to think through problems before responding. Developers building AI coding agents, automated code review pipelines, or multi-step development tools will find it a strong fit. At $1/M input and $2/M output tokens, it offers an accessible price point for agentic, high-throughput use cases.

by xAI | $1/M input | $2/M output

Gemini 3.5 Flash is Google DeepMind's frontier-speed model that combines Flash-tier latency and cost with near-Pro-level reasoning, announced at Google I/O 2026. It processes output 4x faster than comparable frontier models while outperforming Gemini 3.1 Pro on coding and agentic benchmarks — 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, and 84.2% on CharXiv Reasoning. It's purpose-built for agentic workflows: orchestrating multi-step tool use, long-context document analysis, and iterative code generation. With a 1M token context window and full multimodal input support (text, image, audio, video, PDF), it handles complex real-world tasks at scale. At $1.50 per million input tokens and $9.00 per million output tokens, it's the best choice for developers who need frontier intelligence without frontier latency or cost.

by Google | $1.5/M input | $9/M output

Perceptron Mk1 is a vision-language model from Perceptron AI designed for video and embodied reasoning, capable of processing native video at up to 2 frames per second within its 32K context window. It excels at video QA, video summarization, event detection, open-vocabulary object detection and counting, OCR on real-world documents, hand pose estimation, and point-by-example grounding from multimodal prompts. On spatial reasoning benchmarks, Mk1 scores 85.1 on EmbSpatialBench and 72.4 on RefSpatialBench, outperforming frontier models. On VSI-Bench it reaches 88.5, the highest recorded score among compared models. At $0.15 per million input tokens, Mk1 is priced 80–90% below comparable frontier vision-language models, making it a strong choice for developers building video analysis, robotics data curation, or multimodal pipelines at scale.

by Perceptron | $0.15/M input | $1.5/M output

Claude Opus 4.7 Fast is a high-speed configuration of Anthropic's most capable model, delivering up to 2.5x faster output token generation with no reduction in quality or capabilities. It runs the same Opus 4.7 model — which scores 87.6% on SWE-bench Verified (up from 80.8% on Opus 4.6), 94.2% on GPQA Diamond, and 69.4% on Terminal-Bench 2.0 — but optimized for lower latency at premium pricing ($30/$150 per MTok). It supports the full 1M token context window and 128k max output tokens. Fast mode benefits are focused on output tokens per second, not time to first token. It is ideal for latency-sensitive agentic workflows, live coding sessions, and real-time tasks where response speed matters. For cost-sensitive or batch workloads, standard Opus 4.7 offers the same intelligence at lower cost.

by Anthropic | $30/M input | $150/M output

Ring 2.6 1T is a trillion-parameter open-weights reasoning model from InclusionAI (Ant Group), released under the MIT license. It uses a Mixture-of-Experts architecture with approximately 63B active parameters per token and supports a 262K context window with up to 66K output tokens. The model offers adaptive reasoning effort through "high" and "xhigh" modes, letting developers tune thinking depth against token cost based on task complexity. It is purpose-built for agentic workflows, coding agents, tool use, and long-horizon multi-step task execution. Ring 2.6 1T scores 95.83 on AIME 2026, 88.27 on GPQA Diamond, and 87.60 on PinchBench in agent mode — surpassing GPT-5.4 and Gemini 3.1 Pro on that benchmark. A strong pick for developers building autonomous agent systems or complex reasoning pipelines.

by InclusionAI | $0.08/M input | $0.63/M output

Grok 4.3 is xAI's latest flagship reasoning model, designed for agentic workflows, instruction following, and tasks demanding high factual accuracy. It accepts text and image inputs with always-on reasoning that cannot be disabled. The model supports a 1 million token context window with no output token limit, making it well suited for long-document analysis and multi-step agent tasks. Priced at $1.25 per million input tokens and $2.50 per million output tokens, it delivers improved cost-efficiency over its predecessor Grok 4.20 — scoring higher on the Artificial Analysis Intelligence Index while costing roughly 20% less to run. Grok 4.3 showed a major jump in real-world agentic task performance, gaining over 300 Elo points on GDPval-AA versus Grok 4.20. It also scores 98% on τ²-Bench Telecom and 81% on IFBench. A strong pick for developers building cost-sensitive agent systems that need reliable tool use and instruction adherence.

by xAI | $1.25/M input | $2.5/M output

IBM Granite 4.1 8B is a dense, decoder-only language model from IBM, built for enterprise workloads like tool calling, RAG, code generation, summarization, and classification. It supports a 131K-token context window and 12 languages including English, German, Spanish, French, Japanese, and Chinese. Despite its compact size, the 8B model matches or outperforms IBM's previous-generation 32B Mixture-of-Experts model across benchmarks — scoring 69.0 on ArenaHard, 68.3 on BFCL V3 (tool calling), and 92.5 on GSM8K. It implements OpenAI-compatible tool calling and supports fill-in-the-middle for code completion. Its dense architecture makes it straightforward to fine-tune for downstream tasks. Released under the Apache 2.0 license, it's a strong pick for developers who need reliable enterprise capabilities at an efficient parameter count.

by IBM Granite | $0.05/M input | $0.1/M output

Mistral Medium 3.5 is a dense 128-billion-parameter multimodal model from Mistral AI that unifies instruction-following, reasoning, and coding into a single set of weights. It features a 256k-token context window, native function calling, structured JSON output, and vision capabilities via a custom-trained encoder that handles variable image sizes. A per-request reasoning_effort parameter lets you toggle between fast responses and deeper chain-of-thought processing, making the same model suitable for quick chat replies and complex agentic workflows. On benchmarks, it scores 77.6% on SWE-Bench Verified and 91.4% on τ³-Telecom. It replaces Mistral's previous Medium 3.1, Magistral, and Devstral 2 models. Priced at $1.50 per million input tokens and $7.50 per million output tokens, it's a strong fit for developers building tool-calling agents, long-horizon coding tasks, and multi-step automation pipelines.

by Mistral AI | $1.5/M input | $7.5/M output

Nemotron 3 Nano Omni is an open multimodal model from NVIDIA that unifies text, image, video, and audio understanding in a single inference pass. Built on a 30B-parameter hybrid Mamba-Transformer Mixture-of-Experts architecture with only ~3B active parameters per token, it delivers strong reasoning at small-model inference costs. It tops six leaderboards across document intelligence (MMLongBench-Doc, OCRBenchV2), video and audio understanding (WorldSense, DailyOmni, VoiceBench), and achieves the highest throughput of any benchmarked model — open or closed — on MediaPerf's video tasks, with up to 9x higher throughput than comparable open omni models. Designed as a multimodal perception sub-agent in agentic systems, it excels at document reasoning, GUI-based computer use, speech transcription, and audio-video analysis — replacing fragmented multi-model pipelines with a single call. Supports up to 256K context with an optional reasoning mode.

by NVIDIA | $0/M input | $0/M output

Seed 2.0 Mini is ByteDance's smallest and cheapest model in the Seed 2.0 family, available here through BytePlus's hosting (ByteDance's own cloud AI platform, Model Ark). It handles text, image, and video input with tool calling and a 256K context window. ByteDance built it for high-throughput, latency-sensitive workloads rather than deep reasoning. On benchmarks it scores 87.0 on AIME 2025, 67.9 on SWE-Bench Verified, 1644 on Codeforces, and 81.2 on VideoMME, lower than the Pro and Lite models in the same family. At $0.10 per million input tokens and $0.40 per million output tokens, it suits batch classification, content moderation, high-volume chat, and other cost-sensitive API workloads where speed and price matter more than peak capability.

by BytePlus | $0.1/M input | $0.4/M output

Seed 2.0 Lite is ByteDance's mid-tier general-purpose language model, available here through BytePlus's hosting of ByteDance's Model Ark platform. It handles a 256K context window, supports tool calling, and accepts text, image, and video input, with text-only output. On benchmarks ByteDance has reported, it scores 93.0 on AIME 2025, a 2233 Codeforces rating, 81.7 on LiveCodeBench v6, and 73.5 on SWE-Bench Verified. Third-party testing found it ahead of Gemini 3.1 Pro Preview on AIME 2025 and close on MMLU-Pro (87.7 vs 89.8), though behind on SWE-Bench Verified and limited to a smaller context window than Gemini's 1M tokens. As the Lite tier of the Seed 2.0 family, it costs a fraction of the Pro variant per token, suited to high-volume production use like document summarization, customer support, and agent workflows where cost per call matters. This entry is the April 28, 2026 pinned snapshot of the model.

by BytePlus | $0.25/M input | $2/M output

Qwen3.6 Flash is the speed-optimized tier of Alibaba's Qwen3.6 model family, designed for high-throughput, low-latency inference pipelines. It sits alongside Qwen3.6 Max Preview, Plus, and 35B-A3B in the product lineup, targeting use cases where fast response times matter more than peak benchmark scores. Like other Qwen3.6 models, it builds on a hybrid architecture combining linear attention with sparse mixture-of-experts routing. It is best suited for high-volume production workloads such as classification, extraction, summarization, and lightweight agent tasks where latency and cost efficiency are the primary constraints.

by Qwen | $0.17/M input | $0.99/M output

Qwen3.5 Plus is a proprietary hosted model from Alibaba, built on the Qwen3.5-397B-A17B Mixture-of-Experts architecture with 397 billion total parameters and 17 billion active per token. Its headline feature is a 1-million-token native context window — among the largest available via API — making it well suited for processing entire codebases, long documents, or extended multi-turn conversations in a single request. It supports both a deep-thinking mode and an "Auto" mode that adaptively invokes tools like web search and code interpreters. This April 20, 2026 snapshot reflects ongoing improvements to the model since its original February 2026 launch. The Qwen3.5 series demonstrated strong multimodal performance across reasoning, coding, and vision tasks. A solid general-purpose option for developers needing large-context capabilities without migrating to the newer Qwen3.6 line.

by Qwen | $0.3/M input | $1.8/M output

DeepSeek V4 Flash is a Mixture-of-Experts model from DeepSeek, made available here through BytePlus's ModelArk platform. It has 284 billion total parameters with 13 billion activated per token, and supports a 1 million token context window with three reasoning modes (non-think, think high, and think max). In think-max mode, DeepSeek reports 91.6 on LiveCodeBench, 88.1 on GPQA Diamond, and 86.2 on MMLU-Pro, close to the larger V4 Pro's scores on the same benchmarks. Flash reaches most of Pro's reasoning quality at a fraction of the active parameter count. The model is built as the fast, lower-cost member of the V4 family, aimed at coding assistants, chat systems, and agent workflows where response speed and per-token cost matter as much as raw accuracy. Tool calling is supported, so it works in the same agentic pipelines used with DeepSeek's larger models.

by BytePlus | $0.14/M input | $0.28/M output

DeepSeek V4 Pro is a 1.6 trillion-parameter Mixture-of-Experts model from DeepSeek, with 49 billion parameters activated per token. Here it's made available through BytePlus's Model Ark hosting. DeepSeek released it under the MIT license as an open-weight model. The company reports it leads open-source models on math, coding, and STEM reasoning benchmarks. On LiveCodeBench it scores 93.5 in DeepSeek's reported testing, and on SWE-bench Verified it scores 80.6%, matching Gemini 3.1 Pro's reported score on that benchmark. It supports a 1M-token context window and tool calling, which suits long-running coding agents, large codebase review, and other tasks where extended context and multi-step reasoning matter.

by BytePlus | $1.74/M input | $3.48/M output

DeepSeek V4 Flash is a lightweight, efficiency-focused Mixture-of-Experts model from DeepSeek, with 284B total parameters and 13B activated per token. It supports a 1M-token context window and configurable reasoning modes (standard, high, and max thinking effort). Designed as the fast and economical option in the V4 family, Flash delivers reasoning capabilities that closely approach the larger V4 Pro, and performs on par with it on simpler agentic tasks. In its max reasoning mode, it achieves comparable reasoning scores to Pro when given a larger thinking budget. At $0.14/M input and $0.28/M output tokens, it's one of the cheapest frontier-tier models available — well suited for high-throughput workloads like coding assistants, chat systems, and agent pipelines where latency and cost matter most.

by DeepSeek | $0.14/M input | $0.28/M output

DeepSeek V4 Pro is a 1.6T-parameter Mixture-of-Experts model from DeepSeek with 49B parameters activated per token, supporting a 1M-token context window. It is positioned as the strongest open-weight model currently available. V4 Pro leads all open-source models in math, coding, and STEM reasoning. On LiveCodeBench it scores 93.5, ahead of Gemini 3.1 Pro (91.7) and Claude Opus 4.6 (88.8). Its Codeforces rating of 3206 also tops GPT-5.4 (3168). On agentic tool-use benchmarks like MCPAtlas, it reaches near-parity with Opus 4.6. DeepSeek acknowledges it trails GPT-5.4 and Gemini 3.1 Pro overall by roughly 3–6 months of frontier development. Priced at $1.74/M input and $3.48/M output — a fraction of comparable closed-source models — it's a strong pick for complex reasoning, agentic coding, and knowledge-intensive tasks.

by DeepSeek | $1.74/M input | $3.48/M output

GPT-5.5 is OpenAI's newest frontier model for complex professional work — a fully retrained base model in the GPT-5 family. It excels at agentic coding, computer use, knowledge work, and scientific research, and is designed to plan, use tools, and carry multi-step tasks to completion autonomously. It achieves state-of-the-art scores on several benchmarks, including 82.7% on Terminal-Bench 2.0 (agentic coding workflows), 84.9% on GDPval (knowledge work across 44 occupations), and 78.7% on OSWorld-Verified (desktop computer use). It outperforms Claude Opus 4.7 and Gemini 3.1 Pro on most agentic and math benchmarks, though Opus 4.7 still leads on SWE-Bench Pro. Available with a 128,000 max output tokens. Supports image input, structured outputs, function calling, streaming, prompt caching, Batch, distillation, and a full Responses API tool suite including web search, computer use, MCP, hosted shell, and tool search. Defaults to medium reasoning effort (supports none through xhigh).

by OpenAI | $2.5/M input | $15/M output

GPT-5.5 Pro is a version of GPT-5.5 that uses more compute to produce smarter, more precise responses on the hardest problems. It scores 39.6% on FrontierMath Tier 4 (expert-level mathematics) and 43.1% on Humanity's Last Exam (multidisciplinary zero-shot reasoning). Supports function calling, structured outputs, web search, file search, image generation, code interpreter, hosted shell, and MCP. Does not support computer use, apply patch, skills, tool search, or distillation. Shares the same 1,050,000-token context window and 128,000 max output tokens as GPT-5.5. Best suited for legal review, financial modeling, scientific research, and scenarios where first-pass accuracy outweighs cost and latency.

by OpenAI | $30/M input | $180/M output

Ling 2.6 1T is InclusionAI's trillion-parameter flagship non-reasoning model, built by Ant Group's AGI initiative. It uses a Mixture-of-Experts architecture with approximately 50 billion active parameters per token, employing a "fast thinking" approach that reduces token costs to roughly a quarter of comparable models while maintaining top-tier output quality. The model targets advanced coding, complex reasoning, and large-scale agent workflows. It achieves state-of-the-art results on benchmarks like AIME 2025 and SWE-bench Verified, and ranks first among open-source models on ArtifactsBench for front-end code generation. On the Artificial Analysis Intelligence Index, it scores 34 — far above the median of 13 for comparable open-weight non-reasoning models. With a 262K context window and strong tool-use capabilities out of the box, Ling 2.6 1T is a strong fit for developers building autonomous agents or cost-sensitive pipelines that need flagship-level reasoning without a dedicated thinking model.

by InclusionAI | $0.3/M input | $2.5/M output

MiMo V2.5 is a native omnimodal model from Xiaomi that processes text, images, video, and audio within a single architecture and a 1M-token context window. It delivers agentic performance close to its larger sibling, MiMo V2.5 Pro, at roughly half the token cost — scoring 62.3 on ClawEval (general) and 23.8 on ClawEval Multimodal. On video understanding, it reaches 87.7 on Video-MME, competitive with Gemini 3 Pro. Image understanding benchmarks include 81.0 on CharXiv RQ and 77.9 on MMMU-Pro. Priced at $0.40 per million input tokens and $2.00 per million output tokens, MiMo V2.5 is a strong fit for production agent pipelines where you need multimodal perception and reasoning without flagship-tier cost.

by Xiaomi | $0.11/M input | $0.28/M output

MiMo V2.5 Pro is Xiaomi's most capable model, built for complex software engineering, long-horizon agentic tasks, and autonomous multi-step workflows spanning over a thousand tool calls. It scores 57.2 on SWE-bench Pro, 63.8 on ClawEval, and 72.9 on τ3-Bench — placing it alongside Claude Opus 4.6 and GPT-5.4 across most agentic evaluations. Notably, it achieves this while using roughly 40–60% fewer tokens per trajectory than comparable frontier models. The 1M-token context window and 131K max output support entire codebases and extended autonomous sessions. Priced at $1.00 per million input tokens and $3.00 per million output tokens, MiMo V2.5 Pro targets developers building autonomous agents, code-generation pipelines, and complex tool-use workflows where sustained coherence over long contexts is critical.

by Xiaomi | $1/M input | $3/M output

Tencent Hy3 is a 295B-parameter Mixture-of-Experts reasoning model developed by Tencent's Hunyuan team, with only 21B parameters active per query. It supports a 256K-token context window and configurable reasoning levels (disabled, low, high), letting you trade off latency and depth per request. Hy3 is particularly strong on coding and agentic tasks. It scores 74.4% on SWE-bench Verified for real-world bug fixing and 67.1% on BrowseComp for complex web research. Its MoE architecture delivers competitive performance against much larger models — matching Kimi-K2.5 (1T+ parameters) on agent benchmarks at a fraction of the compute cost. Best suited for developers building agentic workflows, code generation pipelines, and multi-step reasoning applications where cost-efficiency matters.

by Tencent | $0.18/M input | $0.6/M output

Qwen3.6 27B is a dense 27-billion-parameter multimodal model from Alibaba's Qwen team, purpose-built for agentic coding and repository-level reasoning. It scores 77.2% on SWE-bench Verified and 59.3% on Terminal-Bench 2.0, outperforming the previous-generation Qwen3.5-397B-A17B across all major coding benchmarks despite being far smaller. It natively supports text, image, and video inputs with a 262K-token context window, extendable to 1M tokens. A standout feature is Thinking Preservation, which retains reasoning traces across conversation turns — reducing redundant computation in multi-step agent loops. The model uses a hybrid attention architecture combining Gated DeltaNet with traditional self-attention. Ideal for developers building coding agents, multi-turn tool-use workflows, or frontend generation pipelines.

by Qwen | $0.6/M input | $3.6/M output

GPT Image 2 is OpenAI's state-of-the-art image generation and editing model, released in April 2026 as the successor to GPT Image 1. It accepts both text and image inputs, enabling generation from prompts as well as editing of existing images. The model is particularly strong at rendering text within images — signs, UI elements, labels, and multi-word strings — which was a persistent weakness in prior OpenAI image models. It also supports non-Latin scripts including Japanese, Korean, Chinese, Hindi, and Bengali. GPT Image 2 outputs up to 2K resolution, supports flexible image sizes, and can generate multiple images from a single prompt. It features a December 2025 knowledge cutoff and built-in reasoning capabilities. Ideal for developers building visual content pipelines, localized marketing assets, infographics, or UI mockups at scale.

by OpenAI | $0.0059/image

Ling 2.6 Flash is a high-efficiency open-weights instruct model from InclusionAI (Ant Group), featuring 104B total parameters with only 7.4B active via a Mixture-of-Experts architecture. It supports a 262K-token context window and is purpose-built for agentic workflows, coding, and document processing. The model scores 26 on the Artificial Analysis Intelligence Index — nearly double the median of 13 among comparable open-weight non-reasoning models, and a 10-point jump over its predecessor Ling-flash-2.0. It also achieves 59.3% on GPQA Diamond. Trained with Agentic Reinforcement Learning, Ling 2.6 Flash is optimized for tool use, terminal operations, and multi-step agent tasks while keeping token consumption notably low. A strong choice for developers building cost-sensitive agent pipelines or high-throughput automation that still demands capable reasoning and code generation.

by InclusionAI | $0.1/M input | $0.3/M output

Qwen3.6 Max Preview is Alibaba's most capable language model to date — a proprietary flagship that claimed the top score on six major coding benchmarks at its April 20, 2026 release. It leads on SWE-bench Pro, Terminal-Bench 2.0, SkillsBench, QwenClawBench, QwenWebBench, and SciCode. The Artificial Analysis Intelligence Index rates it at 52, well above the median for reasoning models in its price tier. It supports a 256K-token context window and is text-only at launch. As a preview release, Alibaba is still actively iterating on the model. Best suited for teams building coding agents, scientific computing tools, or frontend generation systems that need peak benchmark performance.

by Qwen | $1.3/M input | $7.8/M output

Kimi K2.6 is Moonshot AI's latest open-weight multimodal model, built on a 1-trillion-parameter mixture-of-experts architecture with a 256K context window. It excels at agentic coding and long-horizon execution, supporting sustained autonomous workflows with 4,000+ tool calls across languages like Rust, Go, and Python. On key benchmarks, it scores 58.6 on SWE-Bench Pro, 54.0 on HLE with Tools, and 50.0 on Toolathlon — competitive with GPT-5.4 and Claude Opus 4.6 on coding and agent tasks, though trailing them on pure reasoning. The model accepts text, image, and video input, supports both thinking and non-thinking modes, and offers an OpenAI-compatible API. It's a strong pick for developers building multi-step agentic workflows and complex software engineering pipelines.

by Moonshot AI | $0.45/M input | $1.86/M output

Qwen3.6 35B A3B is a sparse Mixture-of-Experts model with 35 billion total parameters but only 3 billion active per token, making it highly efficient for inference. Developed by Alibaba's Qwen team, it scores 73.4% on SWE-bench Verified and 51.5% on Terminal-Bench 2.0 — significantly outperforming dense models like Gemma 4-31B (52.0% on SWE-bench Verified). It natively handles text, image, and video with a 262K-token context window, extendable to 1M tokens. The model supports Thinking Preservation for stable multi-turn reasoning and includes native tool-calling capabilities. Released under Apache 2.0, it was the first open-weight model in the Qwen3.6 family. A strong choice for developers who want frontier-adjacent coding performance at a fraction of the compute cost of larger models.

by Qwen | $0.25/M input | $1.49/M output

Claude Opus 4.7 is Anthropic's most capable generally available model, built for complex reasoning and agentic coding. It offers a step-change improvement in long-horizon agentic work over its predecessor, Opus 4.6, along with strong gains in knowledge work, vision, and file-system-based memory. The model supports a 1M-token context window, 128k max output tokens, and adaptive thinking. It introduces high-resolution image input (up to 2576px / 3.75MP), a new `xhigh` effort level for demanding coding tasks, and task budgets (beta) that let the model self-moderate token usage across an agentic loop. Priced at $5 / $25 per million input/output tokens. Best suited for developers building autonomous agents, multi-step coding workflows, and vision-heavy pipelines where reliability and depth of reasoning matter most.

by Anthropic | $5/M input | $25/M output

GLM-5.1 is a frontier-class reasoning model from Z.ai (formerly Zhipu AI), built as a post-training refinement of GLM-5 with a focus on coding and agentic tasks. It uses a 744B-parameter Mixture-of-Experts architecture with 40B active parameters per token and supports a 200K context window. GLM-5.1 scored 58.4 on SWE-Bench Pro, surpassing GPT-5.4 (57.7) and Claude Opus 4.6 (57.3), and reached 95.3 on AIME 2026. It excels at long-horizon agentic workflows, multi-step tool use, and complex software engineering tasks. The model is text-only — no image or audio input.

by Z.AI | $0.83/M input | $3/M output

Wan 2.7 Text-to-Video is a diffusion-based video generation model from Alibaba, designed to produce cinematic video clips directly from text prompts. It generates native 720p and 1080p video with durations from 2 to 15 seconds, supporting flexible aspect ratios including 16:9, 9:16, and 1:1. A standout feature is optional audio input, which synchronizes character motion and lip movement to a provided audio track during generation. The model responds well to detailed, structured prompts and supports multi-shot narrative control through prompt language alone. It's part of a broader four-model suite that includes image-to-video, reference-to-video, and video editing capabilities. Best suited for marketing content, social media clips, film pre-visualization, and any production pipeline that needs programmatic access to high-quality video generation.

by Wan AI

Gemma 4 26B A4B is a Mixture-of-Experts (MoE) open model from Google DeepMind, built from the same research as Gemini 3. It has 26B total parameters but activates only 3.8B per forward pass, delivering near-31B-dense quality at a fraction of the compute cost. The model supports a 256K token context window, multimodal image and text input, built-in step-by-step reasoning (thinking mode), and native function calling for agentic workflows. It currently ranks #6 among open models on the Arena AI text leaderboard with an estimated LMArena score of 1441 — competitive with models many times its active size. It excels at reasoning, coding, long-context tasks, and structured tool use. It's a strong pick for developers who need high throughput and low latency without sacrificing capability.

by Google | $0.07/M input | $0.34/M output

Qwen 3.6 Plus is Alibaba's flagship large language model, built on a hybrid architecture combining linear attention with sparse mixture-of-experts routing for high throughput and scalability. It's optimized for agentic coding and complex multi-step workflows. On Terminal-Bench 2.0, it scores 61.6, surpassing Claude 4.5 Opus (59.3), while its 78.8 on SWE-bench Verified places it close behind. It also leads on MCPMark (48.2%) for tool-calling reliability. A native multimodal model, it handles text, images, and documents within a 1M-token context window with up to 65K output tokens. Notable features include always-on chain-of-thought reasoning, native function calling, and a preserve_thinking parameter that retains reasoning across multi-turn agent loops. A strong fit for developers building AI coding agents, terminal automation, and tool-using pipelines.

by Qwen | $0.5/M input | $3/M output

Gemma 4 31B is a dense multimodal model from Google DeepMind, built on the same research foundation as Gemini 3. It is the most capable model in the Gemma 4 family, accepting text, image, and video input with a 256K-token context window. It delivers strong benchmark results: 89.2% on AIME 2026, 85.2% on MMLU Pro, 80.0% on LiveCodeBench v6, and 84.3% on GPQA Diamond. On the Arena AI text leaderboard, it ranks as the #3 open model globally, outperforming many models with far higher parameter counts. Gemma 4 31B features native function calling trained into the model, configurable chain-of-thought reasoning, and structured JSON output — making it especially well-suited for agentic workflows, coding tasks, and multi-turn tool use. It supports over 140 languages and serves as a strong foundation for fine-tuning.

by Google | $0.13/M input | $0.38/M output

Trinity Large Thinking is a 398-billion-parameter sparse Mixture-of-Experts reasoning model from Arcee AI, with approximately 13B active parameters per token, post-trained with extended chain-of-thought and agentic reinforcement learning. It generates explicit reasoning traces in thinking blocks before final responses, and its 262K context window accommodates long agentic reasoning chains. Benchmark results include 94.7% on τ²-Bench and 98.2% on LiveCodeBench, placing it at #2 on PinchBench behind only Claude Opus 4.6. Released under Apache 2.0, Trinity Large Thinking is the strongest option in the Trinity family for agentic pipelines, long-horizon planning, complex multi-step coding, and tasks that benefit from transparent reasoning traces.

by Arcee AI | $0.25/M input | $0.9/M output

Grok 4.20 is xAI's flagship large language model, offering a rare combination of low hallucination rates and high throughput at competitive pricing. It achieved a record 78% non-hallucination rate on the Artificial Analysis Omniscience benchmark — the highest of any model tested — making it a strong choice for applications where factual reliability matters more than peak reasoning scores. It scored 78.5% on GPQA Diamond and 87.3% on MATH-500. The model supports a 2M-token context window, text and image inputs, parallel function calling, structured outputs, and built-in web search. Reasoning can be toggled on or off per request via API parameter. At $2 per million input tokens and $6 per million output tokens, it's one of the most affordable frontier models available, with output speeds exceeding 230 tokens per second.

by xAI | $1.25/M input | $2.5/M output

Grok 4.20 Multi-Agent is a variant of xAI's Grok 4.20 purpose-built for orchestrating multiple AI agents that collaborate on complex, multi-step tasks in real time. Rather than relying on a single inference pass, it coordinates parallel agents that independently search, analyze, and cross-reference information before synthesizing a final response. At low or medium reasoning effort it runs 4 agents; at high or extra-high effort it scales to 16. It scored a 68.7 agentic index on Artificial Analysis — among the highest available. The model shares Grok 4.20's 2M-token context window and natively supports web search, X search, and tool orchestration. It generates up to 2M output tokens per response, making it well suited for deep research workflows, multi-source analysis, and long-running agent pipelines.

by xAI | $1.25/M input | $2.5/M output

Veo 3.1 Lite is Google DeepMind's most cost-effective video generation model, built for high-volume applications where per-clip cost is a primary concern. It generates video at the same speed as Veo 3.1 Fast but at less than half the price — starting at $0.05 per second for 720p. The model supports text-to-video and image-to-video with 720p and 1080p output in landscape (16:9) or portrait (9:16), at configurable durations of 4, 6, or 8 seconds. It does not support 4K output, scene extension, or native audio generation — clips are silent by default. Veo 3.1 Lite is ideal for developers building batch video pipelines, social media automation, or interactive tools where cost per generation matters most and audio can be added in post-production.

by Google

Seed 2.0 Code Preview is ByteDance's coding-focused model from the Seed 2.0 family, offered here through BytePlus, ByteDance's cloud AI platform. The 260328 in the id marks a dated snapshot from March 28, 2026, so behavior stays fixed as ByteDance updates the model. It's built for agentic software work such as generating code, debugging, and repository-level tasks like reviewing and revising changes across multiple files, rather than single-line autocomplete. Alongside text, it takes image and video input, so screenshots, design mockups, or short screen recordings can be included directly in a coding request. With a 256K context window, up to 128K output tokens, and tool calling, it fits agent pipelines that need to read a large codebase and act on it. As a preview model, it suits testing and evaluation ahead of a stable release rather than production workloads that need long-term stability.

by BytePlus | $0.5/M input | $3/M output

Seed 2.0 Pro is ByteDance's flagship reasoning and coding model, accessible here through BytePlus's hosting as a dated snapshot (260328, a March 28, 2026 build). It handles text, image, and video input and is built for long-chain reasoning, coding, and agent workflows. ByteDance reports 98.3 on AIME 2025, 87.8 on LiveCodeBench v6, and 76.5% on SWE-Bench Verified, with a Codeforces rating of 3020. On agent benchmarks it reaches 90.4 on tau2-Bench Retail and 77.3 on BrowseComp, and it ranks sixth on the LMArena leaderboard alongside GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro. With a 256K context window, 128K max output, and tool calling, it fits coding agents and multi-step workflows that need a long context and function calling. Pinning this snapshot keeps behavior stable for workloads tuned against a specific build.

by BytePlus | $0.5/M input | $3/M output

KAT-Coder-Pro V2 is the flagship agentic coding model from Kwaipilot (Kuaishou's AI research division), built for enterprise-grade software engineering and SaaS integration. It uses a Mixture-of-Experts architecture with 72B active parameters and offers a 256K token context window. The model achieves a 79.6% solve rate on SWE-Bench Verified, placing it among the top code generation models globally. It scores 44 on the Artificial Analysis Intelligence Index, well above the median of 15 for comparable non-reasoning models in its price tier, and generates output at roughly 109 tokens per second. KAT-Coder-Pro V2 is designed for large-scale production environments, multi-system coordination, and agentic coding workflows. It also supports tool use, function calling, and web aesthetics generation for producing landing pages and presentation decks.

by KwaiPilot | $0.3/M input | $1.2/M output

Reka Edge is a 7B multimodal vision-language model that processes text, image, and video inputs with industry-leading performance in its size class for visual reasoning, object detection, and agentic tool-use. It features a ConvNeXt V2 vision encoder that extracts only 64 tokens per image tile, enabling exceptionally fast and low-latency inference ideal for real-time applications like robotics, automotive, and augmented reality. It demonstrates frontier-level tool-calling abilities and strong temporal video reasoning, outperforming comparable models on benchmarks like MLVU, MMVU, and RefCOCO.

by Reka AI | $0.1/M input | $0.1/M output

GPT-5.4 Nano is the smallest and cheapest model in the GPT-5.4 family, offering a 400k context window at just $0.20/1M input tokens. It excels at classification, data extraction, ranking, and coding sub-agent tasks, outperforming the previous GPT-5 Mini on SWE-Bench Pro (52.4% vs 45.7%). It's ideal for high-volume, low-latency workloads and as a fast sub-agent in multi-model architectures.

by OpenAI | $0.2/M input | $1.25/M output

MiniMax M2.7 is a proprietary reasoning LLM from Chinese AI startup MiniMax, released on March 18, 2026, notable for being one of the first commercial models to actively participate in its own training through autonomous self-evolution loops. It excels at agentic coding workflows with a 56.2% score on SWE-Pro and strong performance in office productivity tasks, scoring the highest ELO (1495) on GDPval-AA among open-source-tier models. It targets developers building complex agent systems and automated workflows.

by MiniMax | $0.25/M input | $1/M output

MiniMax M2.7 Highspeed is a high-throughput, inference-optimized variant of MiniMax M2.7, delivering approximately 100 tokens per second — roughly 66% faster than the standard version. It shares the same model weights and MoE architecture as M2.7, so output quality and reasoning capability are identical; the speed advantage comes entirely from inference-layer routing and batching optimizations. It supports text and image inputs with a 204K context window and features automatic prompt caching and parallel tool calling. Best suited for live coding assistants, autonomous agent pipelines, and interactive workflows where low latency and high throughput matter.

by MiniMax | $0.6/M input | $2.4/M output

Gemini 3.1 Flash-Lite is Google's fastest and most cost-efficient model in the Gemini 3.1 series, designed for high-volume, latency-sensitive workloads. It delivers 2.5x faster time-to-first-token and 45% higher output throughput than Gemini 2.5 Flash, scoring 86.9% on GPQA Diamond and 76.8% on MMMU Pro. An Intelligence Index score of 34 — nearly triple that of 2.5 Flash-Lite — puts it well above prior budget-tier models at the same price point. The model supports text, image, video, audio, and PDF input with a 1M token context window and configurable thinking levels. It's the right choice for developers building translation pipelines, content moderation systems, classification tasks, or any application where throughput and cost matter most.

by Google | $0.25/M input | $1.5/M output

GPT-5.4 Mini is OpenAI's fast, efficient distillation of GPT-5.4, significantly improving over GPT-5 Mini across coding, reasoning, multimodal understanding, and tool use while running 2x faster. It approaches GPT-5.4-level performance on several benchmarks and features a 400k context window.

by OpenAI | $0.75/M input | $4.5/M output

Mistral Small 4 is a 119B-parameter open-source Mixture-of-Experts model (6B active per token) released under Apache 2.0, unifying instruction-following, reasoning, multimodal (text + image), and agentic coding into a single deployment. It features 128 experts, a 256k context window, and configurable reasoning effort that lets developers toggle between fast responses and deep step-by-step reasoning per request. Compared to its predecessor Mistral Small 3, it delivers 40% lower latency and 3x higher throughput while matching or surpassing GPT-OSS 120B on key benchmarks.

by Mistral AI | $0.15/M input | $0.6/M output

GLM-5 Turbo is a foundation model by Z.ai optimized for fast inference and agent-driven workflows, excelling at tool invocation, complex instruction decomposition, and long-chain task execution in OpenClaw scenarios. It is built on top of the GLM-5 architecture (744B parameters, 40B active) with DeepSeek Sparse Attention for reduced deployment cost and up to 205K token context. GLM-5 Turbo supports reasoning/thinking mode and is designed for real-world multi-step agentic tasks including scheduled, persistent, and high-throughput operations.

by Z.AI | $0.79/M input | $3.44/M output

Nemotron 3 Super is NVIDIA's open-weight 120B-parameter hybrid Mamba-Transformer MoE model with only 12B active parameters, designed for running complex multi-agent agentic AI systems at scale. It features a 1-million-token context window to prevent goal drift across long tasks and delivers up to 5x higher throughput than its predecessor. The model excels at reasoning, coding, and tool use.

by NVIDIA | $0.1/M input | $0.5/M output

Seed 2.0 Lite is ByteDance's mid-tier general-purpose LLM that balances strong performance with cost efficiency, scoring 93 on AIME 2025 and 2233 on Codeforces while supporting text, image, and video understanding plus tool-calling capabilities. It serves as the default production-grade model in the Seed 2.0 family, handling roughly 95% of enterprise workloads at about half the cost of the flagship Pro variant. It supports a 256K context window and is positioned as a high-performance alternative for tasks like code review, document processing, information synthesis, and agent-based workflows.

by ByteDance Seed | $0.25/M input | $2/M output

Grok 4.20 (Reasoning) is xAI's deliberate-reasoning variant of the Grok 4.20 model family, designed for tasks that benefit from extended chain-of-thought before producing a final answer. It generates visible thinking traces prior to each response, improving accuracy on complex problems such as multi-step math, code debugging, logical analysis, and scientific reasoning. Reasoning traces are returned in the API response, making them inspectable for debugging and auditing workflows. The model supports a 2M-token context window, text, image, and PDF inputs, parallel function calling, and structured outputs. At $1.25 per million input tokens and $2.50 per million output tokens, it is a cost-effective choice for developers who need the higher answer accuracy of deliberate reasoning without the premium pricing of larger multi-agent variants.

by xAI | $1.25/M input | $2.5/M output

Grok 4.20 (Non-Reasoning) is a fast, direct-response variant of xAI's Grok 4.20 that skips extended chain-of-thought processing to prioritize speed and lower latency. By bypassing the internal reasoning step, it delivers output at roughly 140–230 tokens per second with sub-second time-to-first-token — making it one of the faster frontier models available via API. It shares the same 2-million-token context window as the full Grok 4.20 family and supports text, image, and PDF inputs, tool calling, and native web search. Best suited for high-throughput production workloads where response speed matters more than deep analytical reasoning: summarization, classification, content generation, conversational interfaces, and agentic pipelines requiring fast tool-call decisions.

by xAI | $1.25/M input | $2.5/M output

GPT-5.4 is OpenAI's latest frontier model released on March 5, 2026, designed for complex professional work with a 1.05M token context window, built-in computer-use capabilities, and improved coding from GPT-5.3-Codex. It is 33% less likely to make factual errors per claim compared to GPT-5.2 and scores 83% on OpenAI's GDPval knowledge work benchmark.

by OpenAI | $2.5/M input | $15/M output

GPT-5.4 Pro is a higher-compute version of GPT-5.4 that allocates more reasoning time to produce smarter and more precise answers on complex tasks. It supports reasoning effort levels of medium, high, and xhigh, and shares the same 1.05M token context window as GPT-5.4.

by OpenAI | $30/M input | $180/M output

Mercury 2 is a diffusion-based reasoning language model from Inception Labs that refines all tokens in parallel rather than generating them sequentially, achieving over 1,000 tokens per second — roughly 5x faster than speed-optimized competitors like Claude Haiku and GPT-5 Mini at comparable quality. On reasoning benchmarks, Mercury 2 scores 91.1 on AIME 2025 and 73.6 on GPQA. It also placed second on the Copilot Arena leaderboard for quality while ranking first for speed overall. With a 128K context window, it is purpose-built for latency-sensitive applications — real-time assistants, high-throughput pipelines, and cost-conscious production workloads where reasoning capability matters.

by Inception | $0.25/M input | $0.75/M output

Qwen Image 2.0 is Alibaba's second-generation image foundation model, delivering a major upgrade over the original Qwen Image with a leaner 7B-parameter architecture that outperforms its 20B predecessor across the board. It generates natively at 2048×2048 resolution and unifies text-to-image generation and image editing into a single model — no separate pipelines needed. The model scores 88.32 on DPG-Bench, surpassing FLUX.1 (83.84) and GPT Image 1 (85.15), and ranks #1 on AI Arena's blind human evaluation for both generation and editing. Its headline feature is professional typography rendering: it handles prompts up to 1,000 tokens and can generate complete infographics, PPT slides, posters, and comics with accurate bilingual text layout. Ideal for developers building design-oriented workflows where text accuracy and prompt adherence are critical.

by Qwen | $0.04/image

Qwen Image 2.0 Pro is the highest-fidelity configuration of Alibaba's Qwen Image 2.0, built on the same 7B-parameter architecture but tuned to maximize visual quality over speed. Compared to the standard tier, Pro delivers richer color accuracy, finer detail rendering — visible in textures like hair strands, fabric weaves, and metallic reflections — and stronger adherence to complex, multi-element prompts. Text rendering is also crisper, making it better suited for commercial assets like branded posters and packaging. The standard Qwen Image 2.0 is optimized for fast iteration and prototyping. Pro is where you go for final production renders where every pixel matters. Best for developers building pipelines that need polished, client-ready output from a single API call.

by Qwen | $0.08/image

Gemini 3.1 Flash-Lite Preview is the preview release of Google's fastest and most cost-efficient model in the Gemini 3.1 series, made available through the Gemini API in AI Studio and Vertex AI on March 3, 2026, ahead of the stable version. It scores 86.9% on GPQA Diamond and 76.8% on MMMU Pro, with a 1432 Elo score on the Arena.ai leaderboard. Google reports 2.5x faster time-to-first-token and 45% higher output speed than Gemini 2.5 Flash, and Artificial Analysis measured about 382 output tokens per second. The model supports multimodal input with a 1M token context window and configurable thinking levels. It's built for high-volume tasks like translation, content moderation, classification, and data labeling. As a preview endpoint, its behavior may change before the stable release.

by Google | $0.25/M input | $1.5/M output

Seed 2.0 Lite is ByteDance's mid-tier general-purpose language model, available here through BytePlus's hosting of ByteDance's Model Ark platform. It handles a 256K context window, supports tool calling, and accepts text, image, and video input, with text-only output. On benchmarks ByteDance has reported, it scores 93.0 on AIME 2025, a 2233 Codeforces rating, 81.7 on LiveCodeBench v6, and 73.5 on SWE-Bench Verified. Third-party testing found it ahead of Gemini 3.1 Pro Preview on AIME 2025 and close on MMLU-Pro (87.7 vs 89.8), though behind on SWE-Bench Verified and limited to a smaller context window than Gemini's 1M tokens. As the Lite tier of the Seed 2.0 family, it costs a fraction of the Pro variant per token, suited to high-volume production use like document summarization, customer support, and agent workflows where cost per call matters. This entry is the February 28, 2026 pinned snapshot of the model.

by BytePlus | $0.25/M input | $2/M output

Gemini 3.1 Flash Image (also known as Nano Banana 2) is Google DeepMind's latest state-of-the-art image generation and editing model, combining Pro-level quality with the speed of the Flash architecture. It supports text and image input with up to 1M token context, generates images up to 4K resolution, and features advanced world knowledge, precise text rendering, subject consistency, and web-search grounding.

by Google | $0.067/image

Seed 2.0 Mini is ByteDance's most lightweight and inference-efficient model in the Seed 2.0 family, released in February 2026 and optimized for low-latency, high-concurrency, and cost-sensitive applications. It features a 256K context window, multimodal capabilities (text, image, video), and a unique 4-level reasoning effort system. Despite being the smallest variant, it delivers strong benchmark scores (AIME 2025: 87.0, SWE-Bench: 67.9) at an extremely competitive price of $0.1/M input tokens.

by ByteDance Seed | $0.1/M input | $0.4/M output

Qwen 3.5 Flash is the production-optimized API version of the 35B-A3B model. It features a default 1M token context window, built-in tool/function calling support, and is priced at ~$0.10/M input tokens for low-latency agentic workflows. The '02-23' suffix indicates the February 23, 2026 snapshot/version date.

by Qwen | $0.07/M input | $0.26/M output

Qwen 3.5 Flash is the fast, low-cost tier of the Qwen 3.5 family from Alibaba's Qwen team, served as the hosted API version of the Qwen3.5-35B-A3B model. It uses a hybrid architecture that pairs linear attention with sparse mixture-of-experts routing, which keeps compute close to linear across its 1M token context window. The model supports tool and function calling and accepts image input alongside text, with output up to 64K tokens per request. At $0.10 per million input tokens and $0.40 per million output tokens, it is priced for high-volume work such as classification, extraction, and lightweight agent loops where cost and context length matter more than peak reasoning. Qwen has not published benchmark scores for the Flash tier specifically. This rolling route tracks the latest snapshot; the dated 02-23 snapshot is listed separately.

by Qwen | $0.1/M input | $0.4/M output

GPT-5.3 Codex is OpenAI's most capable agentic coding model, combining frontier coding performance with strong general reasoning and professional knowledge capabilities. It was the first model instrumental in creating itself, having been used to debug its own training and manage its own deployment. It sets state-of-the-art on SWE-Bench Pro and Terminal-Bench while being 25% faster than its predecessor.

by OpenAI | $1.75/M input | $14/M output

Qwen 3.5 122B (10B Active) is Alibaba's largest medium-sized MoE model, activating only 10B of its 122B total parameters per inference pass. It excels at agentic tasks like tool use and multi-step reasoning, leading the Qwen 3.5 lineup on benchmarks such as BFCL-V4 and BrowseComp. It supports 262K native context (extendable to 1M), native multimodal input, and 201 languages under Apache 2.0.

by Qwen | $0.4/M input | $3.2/M output

Qwen 3.5 27B is the only dense (non-MoE) model in the Qwen 3.5 medium series, activating all 27B parameters on every forward pass for maximum per-token reasoning density. It ties GPT-5 mini on SWE-bench Verified at 72.4 and is competitive with Claude Sonnet 4.5 on visual reasoning benchmarks. It runs well on consumer hardware and is open-weight under Apache 2.0.

by Qwen | $0.3/M input | $2.4/M output

Qwen 3.5 35B (3B Active) is a sparse MoE model that activates just 3B of its 35B total parameters, yet outperforms the previous-generation 235B flagship across language, vision, coding, and agent tasks. It uses a hybrid Gated DeltaNet + MoE architecture and can run on GPUs with as little as 8GB VRAM when quantized. It's the base model behind the hosted Qwen 3.5 Flash API.

by Qwen | $0.25/M input | $2/M output

Aion 2.0 is a fine-tuned variant of DeepSeek V3.2, developed by AionLabs and optimized for immersive roleplaying and storytelling. It excels at generating narratives with natural tension, conflict, and dramatic stakes, and handles mature or darker themes with notable nuance. The model offers a 131K-token context window with up to 32K tokens of output, making it well-suited for long-form creative sessions. It supports function calling and streaming. On third-party benchmarks, it has scored 99.5% on general knowledge, 96% on mathematics, and 93.5% on coding tasks. Aion 2.0 is a strong pick for developers building interactive fiction, character-driven chat experiences, or creative writing tools where narrative depth and engagement matter more than raw speed.

by Aion Labs | $0.8/M input | $1.6/M output

Gemini 3.1 Pro is Google's most advanced reasoning model, building on the Gemini 3 series with over double the reasoning performance of its predecessor (77.1% on ARC-AGI-2) and a 1M token context window. It features a three-tier thinking system (low, medium, high) for adjustable reasoning depth and is optimized for agentic workflows, software engineering, and complex problem-solving.

by Google | $2/M input | $12/M output

Claude Sonnet 4.6 is Anthropic's latest mid-tier model released February 2026, delivering near-flagship Opus-level performance in coding, computer use, and agentic tasks at a fraction of the cost ($3/$15 per million tokens). It features a 1M token context window in beta, scores 79.6% on SWE-bench Verified and 72.5% on OSWorld. Developers preferred it over both Sonnet 4.5 (~70% of the time) and even Opus 4.5 (~59%) in real-world coding tests.

by Anthropic | $3/M input | $15/M output

Qwen3.5-Plus is the hosted flagship model in the Qwen3.5 series, available through Alibaba Cloud Model Studio. It offers a 1 million token context window by default and includes built-in tools with adaptive tool use, including web search and code interpreter capabilities. The model supports reasoning mode (chain-of-thought), search, and a fast response mode without extended thinking. It is accessible via an OpenAI-compatible API and can be integrated with third-party coding tools like Claude Code, Cline, and OpenClaw. Qwen3.5-Plus is designed for agentic workflows that combine multimodal reasoning with tool use.

by Qwen | $0.26/M input | $1.56/M output

Qwen3.5 Plus is Alibaba's hosted flagship model in the Qwen3.5 series, built on the Qwen3.5-397B-A17B Mixture-of-Experts architecture with 397 billion total parameters and 17 billion active per token. Its headline feature is a 1-million-token native context window — among the largest available via API — making it well suited for processing entire codebases, long documents, or extended multi-turn conversations in a single request. It supports both a deep-thinking mode and an "Auto" mode that adaptively invokes tools like web search and code interpreters. A solid general-purpose option for developers needing large-context capabilities and agentic workflows that combine multimodal reasoning with tool use.

by Qwen | $0.4/M input | $2.4/M output

Qwen3.5-397B-A17B is an open-weight native vision-language model from Alibaba's Qwen team, released in February 2026. It uses a hybrid architecture combining Gated Delta Networks (linear attention) with a sparse mixture-of-experts design, totaling 397 billion parameters but activating only 17 billion per forward pass for efficient inference. The model delivers strong performance across reasoning, coding, agent tasks, and multimodal understanding, competing with frontier models like GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro. It supports 201 languages and dialects and features a 250k-token vocabulary. Its decoding throughput is reported at 8.6x that of Qwen3-Max under a 32k context length.

by Qwen | $0.6/M input | $3.6/M output

Seed 2.0 Mini is ByteDance's smallest and cheapest model in the Seed 2.0 family, available here through BytePlus's hosting (ByteDance's own cloud AI platform, Model Ark). It handles text, image, and video input with tool calling and a 256K context window. ByteDance built it for high-throughput, latency-sensitive workloads rather than deep reasoning. On benchmarks it scores 87.0 on AIME 2025, 67.9 on SWE-Bench Verified, 1644 on Codeforces, and 81.2 on VideoMME, lower than the Pro and Lite models in the same family. At $0.10 per million input tokens and $0.40 per million output tokens, it suits batch classification, content moderation, high-volume chat, and other cost-sensitive API workloads where speed and price matter more than peak capability.

by BytePlus | $0.1/M input | $0.4/M output

Skylark Pro SC 260215 is a dated snapshot of Skylark Pro, the chat model family ByteDance serves through its BytePlus ModelArk platform. The 260215 identifier corresponds to a February 15, 2026 build. BytePlus positions Skylark Pro as its flagship model for complex question answering, summarization, creative writing, and role-playing. The model line also supports function calling and structured output through the API. This snapshot has a 128,000-token context window and returns up to 12,000 output tokens per response. Pinning a dated snapshot like this one keeps behavior stable across requests, which is useful when a workload has been tuned against a specific build rather than a rolling alias.

by ByteDance | $0.2/M input | $0.8/M output

Seed 2.0 Code is ByteDance's dedicated coding model, released with the Seed 2.0 family in February 2026. It is built for agentic software work rather than autocomplete, with a stated focus on decomposing tasks into plans and revising those plans as work progresses, which suits multi-step feature implementation and bug diagnosis and repair. The model accepts image input alongside text, so it can read design mockups, screenshots, and charts and generate code from them. ByteDance cites high-precision text extraction and chart understanding as part of its multimodal perception. It has a 256K context window and up to 128K output tokens. ByteDance has not published benchmark scores for the Code variant. The same model powers coding in TRAE, ByteDance's AI IDE.

by ByteDance | $0.5/M input | $3/M output

Seed 2.0 Pro is the flagship model in ByteDance's Seed 2.0 series, built for long-chain reasoning, coding, and agent workflows. It accepts text, image, and video input and returns text. On benchmarks it scores 98.3 on AIME 2025, 87.8 on LiveCodeBench v6, and 76.5% on SWE-Bench Verified, with a Codeforces rating of 3020. For agent tasks it reaches 90.4 on tau2-Bench Retail and 77.3 on BrowseComp, and it ranks sixth on the LMArena leaderboard alongside GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro. At $0.50 per million input tokens and $3 per million output tokens with a 256K context window, it is priced well below the frontier models it competes with, a fit for coding agents and complex multi-step workflows.

by ByteDance | $0.5/M input | $3/M output

GLM-5 is Zhipu AI's (Z.ai) fifth-generation flagship open-weight foundation model with 744B total parameters (40B active) in a Mixture of Experts architecture, designed for agentic engineering, complex systems coding, and long-horizon agent tasks. It achieves state-of-the-art performance among open-weight models on coding and agentic benchmarks like SWE-bench Verified and Terminal Bench 2.0, approaching Claude Opus 4.5-level capability.

by Z.AI | $0.57/M input | $1.92/M output

MiniMax M2.5 is a 230B-parameter Mixture-of-Experts model (10B active) from Shanghai-based MiniMax, designed for real-world productivity with state-of-the-art performance in coding (80.2% SWE-Bench Verified), agentic tool use, and search tasks. It rivals top models from Anthropic and OpenAI while costing 1/10th to 1/20th the price, positioning itself as frontier intelligence 'too cheap to meter.' The model excels at full-stack development, office work (Word, Excel, PowerPoint), and autonomous agent workflows.

by MiniMax | $0.15/M input | $0.61/M output

MiniMax M2.5 Highspeed is an optimized variant of M2.5 engineered for ultra-low latency and high-throughput workloads. It delivers the same core intelligence as the standard M2.5—including top-tier coding performance (80.2% SWE-Bench Verified) and strong agentic tool use—through rigorous inference optimization, running at approximately 100 tokens per second. The model is well-suited for latency-sensitive interactive applications, large-scale document automation, and high-frequency agentic pipelines where responsiveness matters as much as quality. Compared to standard M2.5, it trades cost efficiency for significantly faster response times, making it the practical choice when speed is a priority.

by MiniMax | $0.6/M input | $2.4/M output

Qwen3 Max Thinking is Alibaba Cloud's flagship proprietary reasoning model with a 256K context window, featuring test-time scaling and adaptive tool-use capabilities (web search, code interpreter, memory) that allow it to reason iteratively and autonomously. It scores competitively against GPT-5.2 and Gemini 3 Pro on benchmarks like Humanity's Last Exam and HMMT, excelling in math, complex reasoning, and instruction following.

by Qwen | $0.78/M input | $3.9/M output

Claude Opus 4.6 is Anthropic's latest model, released February 2026. It is a powerful model for coding and agentic tasks, with a 200K token context window and a 64K output context window.

by Anthropic | $5/M input | $25/M output

Qwen3-Coder-Next is an open-weight coding model from Alibaba's Qwen team with 80B total parameters but only 3B active per token, designed specifically for coding agents and local development with a 256K context window. It uses a sparse Mixture-of-Experts (MoE) architecture with hybrid attention, trained on 800K executable coding tasks using reinforcement learning to excel at long-horizon reasoning, tool calling, and recovering from execution failures. It achieves performance comparable to models with 10-20x more active parameters on benchmarks like SWE-Bench while maintaining low inference costs.

by Qwen | $0.18/M input | $1.35/M output

Step 3.5 Flash is an open-source reasoning model from StepFun, built on a sparse Mixture-of-Experts (MoE) architecture with 196B total parameters but only 11B active per token. It supports a 256K-token context window and native tool calling. The model is purpose-built for agentic and coding workflows, with generation throughput of 100–300 tokens/sec in typical usage. It scores 74.4% on SWE-bench Verified, 97.3 on AIME 2025, 86.4% on LiveCodeBench-V6, and 88.2 on τ²-Bench. Step 3.5 Flash is a strong choice for developers building AI agents, code assistants, or multi-step reasoning pipelines who need frontier-level intelligence at low per-token cost.

by StepFun | $0.1/M input | $0.3/M output

Seedream 5.0 Lite is ByteDance's cost-optimized entry in the Seedream 5.0 family, released in February 2026 and available here through BytePlus's ModelArk API. Unlike earlier Seedream versions, it runs a multimodal reasoning pass before generating pixels, interpreting spatial relationships and creative intent from incomplete or abstract prompts instead of matching keywords directly. BytePlus positions it for turning structured information, such as diagrams, product compositions, or scientific illustrations, into visuals with a clear layout and hierarchy. On Artificial Analysis's independent Text-to-Image Elo leaderboard it scores 1,200, in the same range as Black Forest Labs' FLUX.2 Dev and below its own Seedream 5.0 Pro sibling (1,283). It suits high-volume, budget-conscious workloads such as presentations and educational or marketing graphics.

by BytePlus | $0.035/image

Dreamina Seedance 2.0 is ByteDance's Seedance 2.0 video model, made available here through BytePlus's hosting. Dreamina is the international edition of ByteDance's Jimeng creation app, and Seedance is the video model behind it, alongside the Seedream image model; this listing is the Seedance 2.0 checkpoint distributed under the Dreamina name. It accepts text, images, video, and audio together as reference input in one generation call, and supports video extension and targeted editing, not just full clip generation. ByteDance describes its motion as physics-compliant and highlights stronger intent understanding and instruction following. Through this API it produces clips up to 15 seconds at 480p, 720p, 1080p, or 4K, 24fps. It is the top tier of the 2.0 family, suited to marketing clips, social content, and multi-shot sequences. The Fast and Mini tiers below it trade resolution ceiling and some quality for lower latency and cost.

by BytePlus

Dreamina Seedance 2.0 Fast is the faster, lower-resolution tier of ByteDance's Seedance 2.0 video family, distributed under the Dreamina brand and available here via BytePlus's hosting. It shares the 2.0 generation's motion engine and multimodal reference input (images, video, and audio alongside a text prompt), but caps output at 720p and 480p rather than the full model's 1080p and 4K ceiling. Duration and frame rate match the standard 2.0 model, up to 15 seconds at 24fps. Trading top-end resolution for speed makes it a fit for high-volume iteration, drafts, and action-heavy or social clips where turnaround matters more than 4K output.

by BytePlus

Solar Pro 3 is a Mixture-of-Experts large language model from Upstage, featuring 102B total parameters with only 12B active per forward pass and a 128K token context window. The model is built for agentic workflows and complex reasoning, trained using Upstage's proprietary SnapPO reinforcement learning framework. It scores 72.3 on Tau2-all (the comprehensive agentic evaluation), roughly doubling its predecessor's 36.0. It also claims 100% schema compliance for structured output generation. Solar Pro 3 is particularly strong in Korean, with robust English and Japanese support — making it a standout choice for multilingual teams operating in East Asian markets. It targets enterprise use cases in domains like finance, healthcare, and legal, where reliable instruction following and structured outputs matter most.

by Upstage AI | $0.15/M input | $0.6/M output

Kimi K2.5 is Moonshot AI's most capable open-source model, a natively multimodal (vision + text) trillion-parameter MoE with 32B active parameters released in January 2026. Built through continual pretraining on ~15 trillion mixed visual and text tokens atop the K2 base, it supports both thinking and instant modes with a 256K context window. It scored 76.8% on SWE-bench Verified, 96.1% on AIME 2025, and 50.2% on Humanity's Last Exam with tools — outperforming Claude Opus 4.5 and GPT-5.2 on the latter. Its standout feature is Agent Swarm, which coordinates up to 100 parallel sub-agents for complex tasks. K2.5 excels at vision-to-code generation, frontend development from screenshots, and large-scale agentic workflows, making it a strong choice for developers building multimodal AI agents.

by Moonshot AI | $0.29/M input | $1.51/M output

MiniMax M2-her is a dialogue-first large language model built for immersive roleplay, character-driven chat, and expressive multi-turn conversations. It stays consistent in tone and personality across conversations and supports rich message roles to learn from example dialogue. This makes it well-suited for storytelling, AI companions, and conversational experiences where natural flow matters.

by MiniMax | $0.3/M input | $1.2/M output

Palmyra X5 is Writer's most advanced enterprise LLM, featuring a 1-million-token context window and adaptive reasoning capabilities designed for agentic AI workflows. The model is purpose-built for orchestrating multi-step agents, with sub-second tool-calling latency (~300ms) and the ability to process a full million-token prompt in roughly 22 seconds. It supports code generation, structured outputs, and over 30 languages. On benchmarks, Palmyra X5 scores 48.7 on BigCodeBench (Full, Instruct), 53% on Longbench v2, and 19.1% on OpenAI's MRCR 8-needle test — close to GPT-4.1's 20.25% on the same evaluation. It's priced at $0.60/M input tokens and $6/M output tokens, positioning it as a cost-efficient alternative for teams building complex, data-heavy agent pipelines at scale.

by Writer | $0.6/M input | $6/M output

GLM 4.7 Flash is designed for speed and efficiency while maintaining strong performance. It features a 200K token context window, making it suitable for processing long documents and generating extended responses.

by Z.AI | $0.06/M input | $0.4/M output

GLM-4.7-FlashX is the fastest inference tier in Z.ai's GLM-4.7 generation, offering the lowest latency in the lineup. It shares the 200K-token context window and core improvements of the 4.7 generation — stronger coding, tool usage, multi-step reasoning, and natural conversational tone — while trading peak capability for maximum speed. The full GLM-4.7 model scores 73.8% on SWE-bench Verified, 84.9% on LiveCodeBench, and 95.7% on AIME 2025. FlashX inherits the same foundational training but is the right pick when response time matters more than squeezing out every point of accuracy. Targets high-throughput coding assistance, real-time agent orchestration, and latency-sensitive chat where the standard GLM-4.7 or GLM-4.7-Flash would be too slow for the concurrency requirements.

by Z.AI | $0.07/M input | $0.4/M output

FLUX.2 Klein 4B is a compact, Apache 2.0 licensed model distilled from the full FLUX.2 architecture, capable of sub-second image generation on consumer GPUs with ~13GB VRAM. It supports text-to-image, image editing, and multi-reference generation in a unified model.

by Black Forest Labs

FLUX.2 Klein 9B is a larger variant of the Klein family built on a 9B flow model with an 8B Qwen3 text embedder, matching or exceeding models 5x its size in quality. It offers higher output diversity than the 4B distilled variant and is ideal for fine-tuning and research.

by Black Forest Labs

Seed 1.8 is ByteDance's agentic multimodal model, available here through BytePlus's Model Ark hosting. It accepts text, image, and video input, and combines search, code, and GUI agent skills in one system. Several thinking modes, from minimal to high, let it adjust reasoning depth to the task. ByteDance reports a score of 67.6 on BrowseComp-en for web search and 87.8 on VideoMME for long-video understanding, and says Seed 1.8 surpasses Gemini 3 Pro on BrowseComp-en. On BytePlus, the model supports a 256,000-token context window, function calling, and video understanding across up to 1,280 extracted frames, useful for agent workflows like itinerary planning, business analysis, and multi-step task execution.

by BytePlus | $0.25/M input | $2/M output

Wan 2.6 Image is a 20-billion-parameter diffusion-based image generation and transformation model developed by Alibaba Cloud. Built on a Multimodal Diffusion Transformer (MMDiT) architecture, it supports text-to-image generation, image-to-image editing, and multi-reference style transfer. The model accepts up to three reference images per request, allowing developers to explicitly control style, subject, and composition by referencing inputs in the prompt (e.g., "image 1" for style, "image 2" for background). It generates outputs up to 2048×2048 pixels across a wide range of aspect ratios. Wan 2.6 Image is particularly strong at localized content generation, with sophisticated understanding of Asian cultural contexts and aesthetics. It's well suited for e-commerce product visualization, brand asset creation, marketing materials, and automated content pipelines where controllability and visual consistency matter more than pure artistic exploration.

by Wan AI | $0.03/image

MiniMax-M2.1 is an enhanced version of M2 with significantly improved multi-language programming capabilities and office scenario support. It features more concise responses, better instruction following, and matches or exceeds Claude Sonnet 4.5 on coding benchmarks while maintaining excellent agent/tool scaffolding generalization.

by MiniMax | $0.27/M input | $0.95/M output

MiniMax M2.1 Highspeed is a latency-optimized variant of MiniMax M2.1, designed for production workloads where inference speed is critical. It delivers approximately 100 tokens per second—roughly 1.7x faster than the standard M2.1—while preserving the same MoE architecture (230B total, 10B active parameters) and output quality. It excels in live coding assistants, autonomous agent loops with chained tool calls, and interactive document analysis with streaming output. Compared to the standard M2.1, it reduces time-to-first-token under concurrent load and supports automatic prompt caching for multi-turn agent pipelines, making it the preferred choice when developer-facing latency matters.

by MiniMax | $0.6/M input | $2.4/M output

Seed 1.6 is a general-purpose multimodal AI model by ByteDance featuring adaptive deep thinking, a 256K context window, and a sparse Mixture-of-Experts architecture with 230B total parameters (23B active per forward pass).

by ByteDance Seed | $0.25/M input | $2/M output

Seed 1.6 Flash is an ultra-fast multimodal model by ByteDance optimized for high-throughput and low-latency inference, supporting text, image, and video inputs with a 256K context window and up to 16K output tokens.

by ByteDance Seed | $0.08/M input | $0.3/M output

GLM-4.7 is an open-weight large language model from Zhipu AI (Z.ai), available here through BytePlus's hosting. It targets coding and agentic workflows, including long-running, multi-step tasks that involve tool use. The model reasons before each response and before each tool call, which Z.ai says improves consistency compared to earlier GLM versions. On benchmarks, GLM-4.7 scores 73.8% on SWE-bench Verified and 95.7% on AIME 2025. Z.ai reports it performs at or above Claude Sonnet 4.5 on SWE-bench Verified, LiveCodeBench v6, and Terminal Bench 2.0. It supports a 200,000 token context window and tool calling, suited for software engineering agents, code generation, and multi-step reasoning tasks.

by BytePlus | $0.6/M input | $2.2/M output

Seed 1.8 is an agentic multimodal model from ByteDance's Seed team that combines search, code, and GUI agent capabilities in a single system. It accepts text, image, and video input and supports three thinking modes that adjust reasoning depth to task complexity. ByteDance reports 67.6 on BrowseComp-en for web search tasks and 87.8 on VideoMME for long-form video understanding, and says the model approaches Gemini 3 Pro on most multimodal tasks while surpassing it on BrowseComp-en. With a 256K context window and tool calling, it fits multi-step agent workflows such as web research, code generation, and GUI automation over screenshots.

by ByteDance | $0.25/M input | $2/M output

Gemini 3 Flash is Google's frontier intelligence model built for speed, combining Pro-grade reasoning with Flash-level latency at a fraction of the cost. It excels at agentic coding, complex analysis, and multimodal understanding with configurable thinking levels.

by Google | $0.5/M input | $3/M output

Ministral 14B is part of the Ministral 3 family, a 14B parameter multimodal model with vision capabilities under Apache 2.0. It offers advanced capabilities for local deployment with instruct, base, and reasoning variants achieving 85% on AIME'25.

by Mistral AI | $0.2/M input | $0.2/M output

Ministral 3B is a compact 3B parameter multimodal model from the Ministral 3 family with vision capabilities. It runs on consumer hardware and edge devices, offering text and image understanding with 256K context in a 3-4GB quantized footprint.

by Mistral AI | $0.04/M input | $0.04/M output

Ministral 8B is an 8B parameter multimodal model offering best-in-class text and vision capabilities for edge deployment. It supports single-GPU operation and provides an optimal balance of performance and efficiency under Apache 2.0.

by Mistral AI | $0.1/M input | $0.1/M output

Seedance 1.5 Pro is ByteDance's audio-visual joint generation model, released in December 2025 and available here through BytePlus's hosting. Unlike earlier Seedance versions, it generates video and audio together in a single pass rather than adding sound afterward, with multi-language lip sync covering Chinese, English, Japanese, Korean, Spanish, and Indonesian, plus Sichuan and Cantonese dialects, and support for cinematic camera moves such as dolly zooms. ByteDance evaluates it on SeedVideoBench 1.5, a benchmark it built with film directors covering prompt following, motion quality, and audio-visual sync, and says the model leads on these measures; no numeric score has been published. Through this API it supports 480p, 720p, and 1080p at up to 12 seconds and 24fps. It targets production use cases with dialogue or sound design built in, such as short dramas, advertising, and animation, where earlier text/image-only Seedance versions would need a separate audio step.

by BytePlus

Nemotron 3 Nano 30B A3B is a 31.6B total parameter (3.2B active) hybrid Mamba-Transformer MoE model trained from scratch by NVIDIA with a 1M token context window. It offers up to 3.3x higher throughput than comparable models and supports configurable reasoning traces for both agentic and conversational tasks.

by NVIDIA | $0.05/M input | $0.2/M output

AutoGLM Phone Multilingual is a 9B-parameter vision-language model from Z.ai purpose-built for autonomous smartphone control. It takes a screenshot of a phone screen, interprets the UI through multimodal perception, and outputs precise actions — taps, swipes, text input — to complete multi-step tasks described in natural language. The multilingual variant extends coverage beyond Chinese-optimized apps to English and other languages, making it suitable for international mobile automation workflows. Its architecture is based on GLM-4.1V-9B-Thinking, and it supports a 66K-token context window. Ideal for developers building mobile testing pipelines, phone-based AI assistants, or cross-app automation agents. Devices are controlled via ADB (Android) or HDC (HarmonyOS), with the model callable through a standard chat completions API.

by Z.AI | $0/M input | $0/M output

GPT-5.2 is OpenAI's flagship model for professional knowledge work and coding, outperforming industry professionals on GDPval across 44 occupations. It excels at spreadsheets, presentations, code, and complex multi-step projects.

by OpenAI | $1.75/M input | $14/M output

GPT-5.2 Chat is the ChatGPT-optimized variant of GPT-5.2 with an August 2025 knowledge cutoff. It provides conversational interactions with the latest world knowledge before requiring web search.

by OpenAI | $1.75/M input | $14/M output

GPT-5.2 Codex is OpenAI's most advanced agentic coding model for professional software engineering and defensive cybersecurity. It achieves state-of-the-art on SWE-Bench Pro with improved long-horizon work through context compaction.

by OpenAI | $1.75/M input | $14/M output

GPT-5.2 Pro is a version of GPT-5.2 that thinks longer to produce smarter and more precise responses for challenging problems. It supports medium, high, and xhigh reasoning effort settings.

by OpenAI | $21/M input | $168/M output

Gemini 3 Pro Image (Nano Banana Pro) is Google's most advanced image generation and editing model built on Gemini 3 Pro, featuring studio-quality output with support for 2K/4K resolution. It excels at accurate text rendering in multiple languages, uses Google Search grounding for real-time data, and employs thinking mode for complex reasoning through prompts.

by Google | $0.1351/image

Devstral 2 is a 123B parameter dense transformer coding model achieving 72.2% on SWE-bench Verified with 256K context. Released under modified MIT license, it's the state-of-the-art open model for code agents, 7x more cost-efficient than Claude Sonnet.

by Mistral AI | $0.4/M input | $2/M output

GLM-4.6V-Flash is a 9B-parameter vision-language model from Z.ai, the lightweight variant of the GLM-4.6V series. It supports a 128K-token context window and processes images, documents, charts, video frames, and text within a single request. Its key differentiator is native multimodal function calling — images and screenshots can be passed directly as tool parameters, and visual tool outputs are consumed in the same reasoning chain. This bridges the gap between visual perception and executable action for multimodal agent workflows. Best for latency-sensitive and cost-conscious applications that need vision-language capabilities: document understanding pipelines, UI-to-code conversion, visual QA, and multimodal agent loops. For maximum accuracy on complex visual reasoning, the full 106B GLM-4.6V model is available.

by Z.AI | $0.1/M input | $0.1/M output

GLM-4.6V-FlashX is the fastest inference tier in Z.ai's GLM-4.6V vision-language model series. Built on the same 9B-parameter architecture as GLM-4.6V-Flash, it shares the 128K-token context window and native multimodal function calling capabilities but is further optimized for throughput and minimal latency. It supports vision input, reasoning, tool use, and structured JSON output — the same feature set as GLM-4.6V-Flash with higher concurrency limits and faster response times. Ideal for high-volume visual processing pipelines where per-request latency is critical: real-time document scanning, automated UI testing at scale, or multimodal chat applications that need vision understanding without waiting on a larger model.

by Z.AI | $0.04/M input | $0.4/M output

Relace Search is an agentic codebase search model that uses 4-12 parallel tool calls (view_file, grep) to explore repositories and return relevant files. It performs multi-step reasoning to produce precise results 4x faster than frontier models, designed to work as a subagent for coding workflows.

by Relace | $1/M input | $3/M output

GPT-5.1 Codex Max is OpenAI's frontier agentic coding model built for long-running, detailed work using context compaction. It's the first model trained to operate across multiple context windows coherently.

by OpenAI | $1.25/M input | $10/M output

Mistral Large 3 is a 675B parameter sparse MoE model (41B active) trained on 3000 H200 GPUs, representing Mistral's frontier open-weight multimodal model. It supports 256K context, native vision, and excels in agentic workflows and enterprise applications.

by Mistral AI | $0.5/M input | $1.5/M output

Amazon Nova 2 Lite is a fast, cost-effective multimodal reasoning model for everyday workloads that processes text, images, and video with a 1M token context window. It features extended thinking with adjustable reasoning intensity (low/medium/high) and built-in tools for web grounding and code execution. Released in December 2025, it excels at document processing, customer service chatbots, and agentic workflows.

by Amazon | $0.3/M input | $2.5/M output

GLM-4.7 is Zhipu AI's latest ~400B flagship released December 2025, optimized for coding with 200K context and 128K output. It scores 73.8% on SWE-bench and 95.7% on AIME 2025.

by Z.AI | $0.4/M input | $1.75/M output

GPT Image 1.5 is OpenAI's latest and most advanced image generation model released in December 2025, offering better instruction following, precise editing, and up to 4x faster generation than GPT Image 1. It maintains details during edits, improves on premature cropping and color bias issues, and is 20% cheaper than its predecessor. This model powers the ChatGPT Images feature and represents the current state-of-the-art in OpenAI's image generation lineup.

by OpenAI | $0.009/image

DeepSeek V3.2 is the December 2025 flagship model featuring DeepSeek Sparse Attention for efficiency and massive reinforcement learning post-training, achieving GPT-5-level performance. It's the first DeepSeek model to integrate thinking directly into tool-use and excels at agentic AI tasks.

by DeepSeek | $0.26/M input | $0.38/M output

DeepSeek V3.2 is a mixture-of-experts large language model from DeepSeek, made available here through BytePlus's Model Ark hosting. It uses DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism that lowers compute cost while preserving quality on long-context tasks, and it supports tool use in both thinking and non-thinking modes. DeepSeek describes its reasoning performance as in the same class as GPT-5, trained with a large-scale agentic data synthesis pipeline spanning over 1,800 environments and more than 85,000 complex instructions to strengthen tool-use in agentic settings. It's suited for reasoning-heavy work such as math and coding, and for agentic applications that call tools as part of multi-step tasks, while balancing inference speed against the deeper reasoning of DeepSeek's higher-compute Speciale variant.

by BytePlus | $0.28/M input | $0.42/M output

Seedream 4.5 is ByteDance's follow-up to Seedream 4.0, announced in December 2025 and available here through BytePlus's ModelArk API. It keeps the same unified generation-and-editing architecture and scales it up rather than introducing a new one. BytePlus reports improvements in editing consistency, including better preservation of subject detail, lighting, and color tone, along with sharper small-text and facial rendering and steadier multi-image fusion across up to 10 reference images. 4K output is included at no extra charge over lower resolutions. On Artificial Analysis's independent Text-to-Image Elo leaderboard, Seedream 4.5 scores 1,207, behind Seedream 4.0's 1,225 on that measure. It targets production use cases such as e-commerce product imagery, marketing campaigns, and game or film pre-production art where consistent editing matters more than one-shot generation.

by BytePlus | $0.04/image

OLMo 3 32B Think is a fully open reasoning model from the Allen Institute for AI (AI2), and the first fully open 32B thinking model to expose intermediate chain-of-thought reasoning traces. Trained on multi-step math, code, and general problem-solving tasks using a thinking SFT, DPO, and RLVR training flow, it is the strongest fully open reasoning model at the 32B scale — narrowing the gap to open-weight models like Qwen 3-32B-Think while trained on 6x fewer tokens. All training data, code, weights, and checkpoints are publicly available under Apache 2.0. Best suited for complex multi-step reasoning and mathematical problem solving via API.

by Allen AI | $0.15/M input | $0.5/M output

Grok 4.1 Fast is xAI's best tool-calling model released November 2025, featuring a 2M context window and halved hallucination rates versus Grok 4 Fast. It comes in reasoning and non-reasoning modes and is optimized for agentic workflows with native support for web search, X search, and code execution.

by xAI | $0.2/M input | $0.5/M output

Grok 4.1 Fast Non-Reasoning is the low-latency, non-reasoning variant of Grok 4.1 Fast that skips extended chain-of-thought for speed-critical applications. It shares the same model weights and 2M context window as Grok 4.1 Fast but delivers instant responses without deliberation overhead, ideal for real-time customer support and streaming interactions.

by xAI | $0.2/M input | $0.5/M output

Grok 4.1 Fast (Reasoning) is xAI's reasoning-enabled variant of Grok 4.1 Fast, designed for tasks that benefit from extended chain-of-thought deliberation within a fast, agentic framework. It shares the same model weights and 2M token context window as Grok 4.1 Fast but activates multi-step reasoning, making it well suited for complex problem solving, multi-hop analysis, and tasks where accuracy matters more than raw throughput. Reasoning effort is configurable via the API (low, medium, or high) to balance cost and quality. The model supports text and image inputs, native tool use including web search and code execution, and parallel function calling. It cuts hallucination rates in half compared to Grok 4 Fast and is optimized for agentic workflows in enterprise domains such as customer support, finance, and research.

by xAI | $0.2/M input | $0.5/M output

GPT-5.1 is OpenAI's model that dynamically adapts reasoning time based on task complexity, making it faster and more token-efficient on simpler tasks. It features 8 customizable personalities and supports multimodal inputs.

by OpenAI | $1.25/M input | $10/M output

GPT-5.1 Chat is the conversational variant of GPT-5.1 used in ChatGPT with a warmer personality by default. It's available as gpt-5.1-chat-latest in the API for non-reasoning chat interactions.

by OpenAI | $1.25/M input | $10/M output

GPT-5.1 Codex is a version of GPT-5.1 optimized for agentic coding tasks in Codex or similar environments. It's designed for long-running coding workflows with enhanced code generation capabilities.

by OpenAI | $1.25/M input | $10/M output

GPT-5.1 Codex Mini is a smaller, more cost-effective version of GPT-5.1 Codex providing approximately 4x more usage within subscription limits. It balances coding capability with efficiency.

by OpenAI | $0.25/M input | $2/M output

Kimi K2 Thinking is Moonshot AI's reasoning-enhanced variant of Kimi K2, trained to interleave step-by-step chain-of-thought with dynamic tool calls. It supports up to 200–300 sequential tool calls without drift, enabling deep autonomous research, coding, and analysis workflows. It achieves 71.3% on SWE-bench Verified, 44.9% on Humanity's Last Exam (with tools), 60.2% on BrowseComp, and 99.1% on AIME 2025 (with Python) — placing it among the top open-source thinking models. It uses native INT4 quantization and a 256K context window. K2 Thinking is designed for complex, multi-step tasks where extended reasoning and sustained tool orchestration matter more than low-latency responses.

by Moonshot AI | $0.6/M input | $2.5/M output

Claude Opus 4.5 was released in November 2025. It sets the standard for production code, sophisticated agents, and complex enterprise tasks—scoring higher than Anthropic's own engineering candidates on technical tests.

by Anthropic | $5/M input | $25/M output

Amazon Nova Premier is the most capable multimodal model in the Nova family, designed for complex reasoning tasks requiring the highest accuracy. It processes text, images, and video with advanced understanding capabilities and serves as the best teacher model for distilling custom variants of smaller Nova models. Best suited for sophisticated enterprise applications demanding top-tier intelligence.

by Amazon | $2.5/M input | $12.5/M output

Sonar Pro Search is Perplexity's most advanced agentic search system, available exclusively via OpenRouter API, adding autonomous multi-step reasoning to Sonar Pro. Instead of single query+synthesis, it plans and executes entire research workflows using tools, making it ideal for deeper reasoning and analysis. This model powers Perplexity's Pro Search mode on their consumer platform.

by Perplexity | $3/M input | $15/M output

Voxtral Small 24B is an open-source speech understanding model built on Mistral Small 3.1 under Apache 2.0. It handles transcription, translation, Q&A, and summarization directly from audio in 8+ languages with 32K token context.

by Mistral AI | $0.1/M input | $0.3/M output

GPT-OSS Safeguard 20B is a safety-focused variant of the 20B open-weight model with additional content moderation capabilities. It includes enhanced safeguards for responsible deployment.

by OpenAI | $0.08/M input | $0.3/M output

Granite 4.0 Micro is a 3B-parameter dense language model from IBM, built on a conventional transformer architecture and optimized for low-latency, cost-efficient workloads. Despite its compact size, it significantly outperforms its predecessor Granite 3.3 8B across the board — a model more than twice its size. It scores 16 on the Artificial Analysis Intelligence Index, placing ahead of Gemma 3 4B (15). In RAG benchmarks, it outperforms much larger models including Llama 3.3 70B and Qwen3 8B. The model natively supports tool calling, function calling, multilingual generation, fill-in-the-middle code completion, RAG, and structured JSON output, with a 128K token context window. It's a strong fit for agentic sub-tasks, API orchestration, and scenarios where speed and cost matter more than peak reasoning power.

by IBM Granite | $0.02/M input | $0.11/M output

DeepSeek OCR is a vision model from DeepSeek, released in October 2025, built for optical character recognition and document parsing. It takes images as input through a chat-style API and returns text, including documents converted to markdown. Its core idea is optical context compression, representing text as a small number of vision tokens and decoding them back to text. In DeepSeek's paper it reaches 97% decoding precision at compression ratios under 10x, and about 60% at 20x. On OmniDocBench it outperforms GOT-OCR 2.0 while using only 100 vision tokens per page (versus 256), and beats MinerU 2.0 with fewer than 800 vision tokens (versus 6,000+). It fits document-heavy workloads: parsing PDFs and scans, extracting text and tables, and generating training data for other models.

by DeepSeek | $0.04/M input | $0.08/M output

Qwen3-VL Flash is the fast, low-cost hosted vision-language model in Alibaba Cloud's Qwen3-VL series, served on Model Studio as the cheaper tier below Qwen3-VL Plus. It accepts text, image, and video input (video files up to one hour or 2 GB per request) and covers the Qwen3-VL task set: OCR in 32 languages, document and chart parsing, spatial reasoning, and GUI-based visual agent tasks. Alibaba reports it beats the smaller Qwen3-VL 30B A3B in both accuracy and response speed. Like Qwen3-VL Plus, it is a hybrid thinking model, answering directly by default with an optional thinking mode via the enable_thinking parameter. With a 256K context window, 32K max output, and function calling support, it fits high-volume visual workloads (screenshot parsing, receipt OCR, video tagging) where per-request cost matters more than peak accuracy.

by Qwen | $0.05/M input | $0.4/M output

Veo 3.1 is Google DeepMind's flagship AI video generation model, offering the highest quality output in the Veo family. It generates up to 4K resolution video with natively synchronized audio — including dialogue, sound effects, and ambient noise — all produced in a single joint diffusion process. The model supports text-to-video, image-to-video, reference image guidance (up to 3 images), and frame-to-frame generation. Clips are 8 seconds at base, extendable to over 60 seconds via scene chaining. Both 16:9 and native 9:16 aspect ratios are supported. Lip-sync accuracy sits under 120ms. Veo 3.1 achieved top human-preference scores on MovieGenBench for prompt adherence, visual quality, and audio sync, and state-of-the-art results on VBench I2V. It was the first major AI video model to support true 4K output. Best suited for high-fidelity creative and production work where quality is the priority over speed or cost.

by Google

Veo 3.1 Fast is the speed-optimized variant of Google DeepMind's Veo 3.1 video model, generating output roughly twice as fast as the standard version with only a minor quality trade-off. Independent testing shows quality differences of 1–8% depending on scene complexity — negligible for most use cases. It retains the full feature set of the standard model: native audio generation, text-to-video, image-to-video, reference images, frame-to-frame generation, and support for 720p, 1080p, and 4K resolutions. An 8-second 720p clip typically completes in 30–45 seconds. At roughly one-fifth the per-second cost of the standard model, Veo 3.1 Fast is well suited for rapid prototyping, iterative prompt testing, and production workflows where turnaround time and budget matter more than maximum fidelity.

by Google

Seedance 1.0 Pro Fast is ByteDance's speed-optimized variant of Seedance 1.0 Pro, offered here through BytePlus's hosting. BytePlus documentation describes it as inheriting Seedance 1.0 Pro's core strengths, including deep instruction parsing, multi-shot storytelling with natural transitions, and detail rendering for faces and hair, while running about 3x faster and at roughly 72% lower output cost than the standard Pro model. Through this API it supports 480p, 720p, and 1080p at up to 12 seconds and 24fps, the same resolution and duration range as Seedance 1.0 Pro. It fits workloads where volume and turnaround matter, such as rapid iteration and high-throughput content pipelines, with Seedance 1.0 Pro itself the better choice when quality matters more than speed.

by BytePlus

Claude Haiku 4.5 is Anthropic's small, fast model released October 2025, optimized for low latency and cost. Despite being the cheapest option ($1/$5 per million tokens), it matches Sonnet 4 on coding benchmarks (73.3% SWE-bench).

by Anthropic | $1/M input | $5/M output

Qwen3 VL 8B Thinking is the reasoning-enhanced compact vision model for complex visual analysis requiring step-by-step reasoning with efficient resource usage.

by Qwen | $0.18/M input | $2.1/M output

Qwen3 VL 30B A3B Instruct is an efficient vision-language MoE model offering strong image/video understanding with 3B active parameters and 256K context support.

by Qwen | $0.13/M input | $0.52/M output

Qwen3 VL 30B A3B Thinking is the reasoning-enhanced vision-language variant optimized for complex visual reasoning tasks with extended thinking capabilities.

by Qwen | $0.2/M input | $2.4/M output

GPT-5 Pro is an enhanced version of GPT-5 that thinks longer using parallel test-time compute to provide the highest quality answers. It replaces o3-pro for complex enterprise and research tasks.

by OpenAI | $15/M input | $120/M output

GPT Image 1 Mini is OpenAI's cost-optimized image generation model released in October 2025, offering the same capabilities as GPT Image 1 at approximately 80% lower cost. It's designed for high-throughput production use cases where cost and latency are priorities over peak image fidelity. The model trades some fine detail and photorealism for significantly reduced pricing.

by OpenAI | $0.005/image

GLM-4.6 is Zhipu AI's 355B-parameter (32B active) flagship text model with 200K context, excelling at coding, agentic workflows, and search tasks. It's 15% more token-efficient than GLM-4.5 and ranks as the #1 domestic model in China.

by Z.AI | $0.43/M input | $1.74/M output

DeepSeek V3.2-Exp is the September 2025 experimental predecessor to V3.2, introducing DeepSeek Sparse Attention architecture through continued training on V3.1-Terminus. It served as a testing ground for the sparse attention innovations later refined in V3.2.

by DeepSeek | $0.26/M input | $0.38/M output

Claude Sonnet 4.5 is Anthropic's most capable model for agents and computer use, released September 2025. It can maintain focus for 30+ hours on complex tasks, supports a 1M token context window (beta), and is described as their "most aligned frontier model."

by Anthropic | $3/M input | $15/M output

Cydonia 24B v4.1 is a 24-billion parameter uncensored creative writing model based on Mistral Small 3.2, optimized for roleplay, storytelling, and long-form narratives with a 131K token context window. It excels at character consistency, descriptive prose without being overly flowery, and maintains good recall and prompt adherence. The model also performs well for coding and instruction-following tasks.

Relace Apply 3 is a specialized code-patching model that merges AI-generated code edits into existing source files at up to 10,000 tokens per second. It supports a 256K context window and works with diffs from models like Claude and GPT-4, making code integration fast and reliable.

by Relace | $0.85/M input | $1.25/M output

Qwen3 Max is the most powerful Qwen3 API model with SOTA agent programming and tool usage capabilities. It features non-thinking mode optimized for complex agent scenarios.

by Qwen | $1.2/M input | $6/M output

Qwen3 VL 235B A22B Thinking is the reasoning-enhanced vision-language model excelling at visual math, detail analysis, and causal reasoning with extended chain-of-thought processing.

by Qwen | $0.5/M input | $2.5/M output

Qwen3-VL Plus is Alibaba Cloud's hosted vision-language API model in the Qwen3-VL series, offering strong multimodal understanding without requiring self-hosted infrastructure. It handles a wide range of visual tasks including document parsing, chart analysis, OCR, image reasoning, and GUI interaction for PC and mobile interfaces. With a 262K token context window, it is well suited for processing lengthy documents, multi-page PDFs, and extended visual conversations in a single request. The model supports structured output and tool calling, making it a practical choice for developers building document intelligence pipelines, visual agents, and multimodal data extraction workflows via the OpenAI-compatible Alibaba Cloud Model Studio API.

by Qwen | $0.2/M input | $1.6/M output

Qwen3 Omni 30B A3B Instruct is a natively multimodal Mixture-of-Experts model from Alibaba's Qwen team, with 30 billion total parameters and about 3 billion active per token. It accepts text, image, audio, and video input, and uses a Thinker-Talker design that can stream speech as well as text, with a reported end-to-end first-packet latency of 234 ms. The Qwen team reports state-of-the-art results on 22 of 36 audio and audio-visual benchmarks, and open-source state-of-the-art on 32, with speech recognition and voice conversation performance comparable to Gemini 2.5 Pro and ahead of GPT-4o-Transcribe. It covers 119 text languages, 19 speech input languages, and 10 speech output languages. This Instruct variant answers directly without chain-of-thought (a separate Thinking variant does extended reasoning). It supports function calling and fits voice assistants, transcription, audio analysis, and multimodal chat.

by Qwen | $0.25/M input | $0.97/M output

Qwen3 Omni 30B A3B Thinking is the chain-of-thought reasoning variant of Qwen3-Omni, a natively multimodal Mixture-of-Experts model from Alibaba's Qwen team with 30 billion total parameters and about 3 billion active per token. It accepts text, image, audio, and video input and understands text in 119 languages and speech in 19. Unlike the Instruct variant, which can stream speech replies, the Thinking variant contains only the Thinker component and outputs text, producing explicit reasoning before the answer. Reported scores include 73.7 on AIME25, 73.1 on GPQA, and 88.8 on MMLU-Redux for text reasoning, 62.9 on MathVision and 69.7 on Video-MME for vision, and 75.8 on DailyOmni for audio-visual understanding. It fits harder cross-modal tasks, such as math problems presented in images or reasoning over audio and video content.

by Qwen | $0.25/M input | $0.97/M output

DeepSeek V3.1-Terminus is the September 2025 refined update to V3.1, addressing user-reported issues like language mixing and improving Code Agent and Search Agent capabilities. It represents the final, most stable version of the V3 architecture before V3.2.

by DeepSeek | $0.27/M input | $0.95/M output

Grok 4 Fast is an optimized variant delivering performance similar to Grok 4 but using 40% fewer thinking tokens with a massive 2 million token context window. It offers up to 64x cost reduction versus early frontier models like OpenAI's o3.

by xAI | $0.2/M input | $0.5/M output

Grok 4 Fast Non-Reasoning is the speed-optimized, non-reasoning variant of Grok 4 Fast that bypasses extended chain-of-thought for instant responses. It uses the same unified architecture with a 2M context window but skips deliberation, delivering up to 342 tokens/second throughput for high-volume, latency-sensitive applications.

by xAI | $0.2/M input | $0.5/M output

Tongyi DeepResearch 30B A3B is an agentic deep-research model from Alibaba's Tongyi lab, built on Qwen3-30B-A3B with a mixture-of-experts design that activates 3.3B of its 30.5B parameters per token. It is trained for long-horizon web research, multi-step information seeking, and report synthesis rather than general chat. It scores 32.9 on Humanity's Last Exam, 43.4 on BrowseComp, 46.7 on BrowseComp-ZH, 75.0 on xbench-DeepSearch, and 90.6 on FRAMES, and the Tongyi team reports performance on par with OpenAI's Deep Research across these benchmarks. Training combines agentic continual pre-training, supervised fine-tuning, and reinforcement learning on fully synthetic data. It runs in two inference modes, a native ReAct loop that needs no prompt engineering and a Heavy mode based on the IterResearch paradigm, which rebuilds a streamlined workspace each research round and can run agents in parallel. A fit for developers building autonomous research agents and deep search pipelines.

by Qwen | $0.09/M input | $0.45/M output

Qwen3-Omni Flash is a fast, cost-efficient omni-modal model from Alibaba's Qwen3 series, designed for real-time multimodal applications. As a member of the Qwen3-Omni family, it ingests text, images, audio, and video in a single end-to-end architecture — no separate pipelines or modality-switching required. It produces text responses and supports low-latency streaming, making it well suited for voice assistants, live audio/video analysis, and cost-sensitive production workloads. The Flash tier prioritizes speed and throughput over the maximum capability of the full Qwen3-Omni model, with a 65K context window and 16K output limit optimized for shorter media clips and high-volume inference. Developers building real-time assistants, transcription tools, or multimodal agents who need broad input coverage at a lower cost point will find it a practical choice.

by Qwen | $0.43/M input | $1.66/M output

Seed 1.6 is a multimodal reasoning model from ByteDance's Seed team, accessible here through BytePlus's Model Ark hosting. It reads text, images, and video across a 256,000-token context window, using a sparse Mixture-of-Experts design with 230 billion total parameters and 23 billion active per forward pass. The model offers three reasoning modes, thinking, non-thinking, and auto, plus an adaptive chain-of-thought mechanism that decides whether to reason step by step based on prompt complexity, rather than always running deep reasoning. ByteDance reports Seed 1.6-Thinking scored 683/750 in the humanities section of China's 2025 Gaokao exam (ranked first among models tested) and 648/750 in science (ranked second), plus 100% accuracy on the math section of India's JEE Advanced exam. ByteDance compared these results against Gemini 2.5 Pro, DeepSeek-R1, Claude Sonnet 4, and OpenAI o3. Tool calling is supported, fitting agent and document workflows that combine multimodal understanding with API-driven actions.

by BytePlus | $0.25/M input | $2/M output

Qwen Plus (2025-07-28) is a snapshot version of Qwen Plus from July 2025, offering consistent behavior and performance for production deployments requiring version stability.

by Qwen | $0.26/M input | $0.78/M output

Qwen Plus (2025-07-28) Thinking is the reasoning-enhanced version that uses chain-of-thought processing for complex problems, providing step-by-step reasoning before delivering answers.

by Qwen | $0.26/M input | $0.78/M output

Nemotron Nano 9B V2 is a 9B parameter hybrid Mamba-Transformer reasoning model from NVIDIA with a 128K context window. Most self-attention layers are replaced with Mamba-2 layers, which NVIDIA reports gives up to 6x higher inference throughput than Qwen3-8B when generating long reasoning traces. On NVIDIA's published evaluations it scores 72.1% on AIME25, 97.8% on MATH500, 64.0% on GPQA, and 71.1% on LiveCodeBench, ahead of Qwen3-8B on each. It supports tool calling and covers English, German, Spanish, French, Italian, and Japanese. Reasoning can be toggled on or off, and a runtime thinking budget caps how many reasoning tokens the model uses before answering, which NVIDIA says can cut reasoning costs by up to 60%. It fits chatbots with strict latency targets, tool-using copilots, and RAG pipelines that need predictable step times.

by NVIDIA | $0.04/M input | $0.16/M output

Kimi K2 0905 is Moonshot AI's September 2025 update to the original Kimi K2, delivering enhanced coding performance and improved tool-calling reliability. It shares the same 1-trillion-parameter MoE architecture with 32B active parameters but doubles the context window from 128K to 256K tokens. Key improvements include stronger frontend development capabilities — producing cleaner, more polished UI code for frameworks like React, Vue, and Angular — along with better integration across popular agent scaffolds. It scored 53.7% Pass@1 on LiveCodeBench. This version is ideal for developers who want K2's agentic strengths with improved real-world coding quality and longer context support for large codebases.

by Moonshot AI | $0.6/M input | $2.5/M output

Qwen3 Next 80B A3B Instruct is an innovative MoE model with hybrid attention (Gated DeltaNet + Gated Attention), achieving 10x inference throughput for 32K+ contexts while matching Qwen3-235B performance.

by Qwen | $0.5/M input | $2/M output

Qwen3 Next 80B A3B Thinking is the reasoning-enhanced variant outperforming Gemini-2.5-Flash-Thinking on complex reasoning tasks with hybrid attention and multi-token prediction.

by Qwen | $0.5/M input | $6/M output

Magistral Medium is Mistral's enterprise reasoning model with chain-of-thought capabilities, scoring 73.6% on AIME2024 (90% with majority voting). It excels in multilingual step-by-step reasoning for legal, financial, and scientific applications.

by Mistral AI | $2/M input | $5/M output

Magistral Small is a 24B parameter open-source reasoning model under Apache 2.0, achieving 70.7% on AIME2024. It provides traceable, multilingual chain-of-thought reasoning in English, French, Spanish, German, Italian, Arabic, Russian, and Chinese.

by Mistral AI | $0.5/M input | $1.5/M output

MiniMax-M2 is a compact MoE model (230B total, 10B active parameters) optimized for coding and agentic workflows with a 128K context window. It ranks #1 among open-source models for tool use and agent tasks, delivering elite performance in multi-step development workflows at 8% the cost of comparable models.

by MiniMax | $0.26/M input | $1/M output

Hermes 4 14B is the smallest model in Nous Research's Hermes 4 family, a fine-tune of Qwen3-14B trained on a post-training corpus of roughly 60B tokens built around verified reasoning traces. It is a hybrid reasoning model. It answers directly by default and can be switched into a reasoning mode that works through explicit think segments before the final answer, with training aimed at keeping those reasoning traces from running overlong. It is trained for function calling and tool use within a single assistant turn, and for structured outputs, including emitting schema-valid JSON and repairing malformed objects. In line with the Hermes series' focus on steerability and neutral alignment, Nous Research reports state-of-the-art results for Hermes 4 on its RefusalBench test, meaning fewer refused requests than comparable models.

by Nous Research | $0.01/M input | $0.05/M output

Codestral is Mistral's cutting-edge code generation model supporting 80+ programming languages with optimized low-latency performance. It specializes in fill-in-the-middle completion, code correction, and test generation with 2.5x faster performance than its predecessor.

by Mistral AI | $0.3/M input | $0.9/M output

Grok Code Fast 1 is a speedy, economical reasoning model built from scratch specifically for agentic coding workflows, released August 2025. It excels at TypeScript, Python, Java, Rust, C++, and Go with a 256K context window and ~92 tokens/second throughput.

by xAI | $0.2/M input | $1.5/M output

Qwen3 30B A3B Thinking (2507) is the reasoning-enhanced variant optimized for complex problem-solving with extended chain-of-thought processing at high parameter efficiency.

by Qwen | $0.08/M input | $0.4/M output

Seedream 4.0 is ByteDance's unified text-to-image and image-editing model, available here through BytePlus's hosted API. Released in September 2025, it replaced Seedream 3.0's separate generation and editing pipelines with a single architecture that handles both. It generates images up to 4K resolution, more than 10 times faster than Seedream 3.0, and accepts up to 10 reference images at once while keeping characters and styles consistent across a series of outputs. ByteDance reports that on its own MagicArena Elo evaluation, Seedream 4.0 surpassed Google's Gemini 2.5 Flash Image, known as Nano Banana, in single-image editing. On Artificial Analysis's independent Text-to-Image Elo leaderboard it scores 1,225, ahead of Black Forest Labs' FLUX.2 Dev (1,200) though behind newer flagship models. It fits multi-image storytelling and editing workflows where character and style consistency across a sequence of images matters.

by BytePlus | $0.03/image

Hermes 4 405B is a frontier hybrid-mode reasoning model based on Llama-3.1-405B, trained on a 60B token dataset with verified reasoning traces. It features toggleable deep reasoning via think tags, massive improvements in math, code, STEM, and logic, and achieves state-of-the-art on RefusalBench for reduced censorship.

by Nous Research | $0.09/M input | $0.37/M output

Hermes 4 70B is a hybrid reasoning model based on Llama-3.1-70B with toggleable deep thinking mode using think tags. It offers major improvements in math, code, STEM, logic, and creative writing while supporting JSON schema adherence, function calling, and reduced refusal rates compared to other models.

by Nous Research | $0.05/M input | $0.2/M output

Gemini 2.5 Flash Image Preview is Google's image generation and editing model, widely known by its codename Nano Banana. It generates and edits images through the same chat-style API as other Gemini models, taking text and images as input and returning images alongside text. It is known for editing images with natural-language instructions, keeping a character's appearance consistent across multiple edits, and blending several input images into one coherent scene. At release in August 2025 it ranked #1 on LMArena's leaderboards for both text-to-image generation and image editing. Output is billed per token at $30 per million. Each generated image is 1290 output tokens, about $0.039 per image. The model is in preview, so the ID and behavior may change. Pick it for conversational, multi-turn image generation and editing rather than plain text chat.

by Google | $0.3/M input | $30/M output

DeepSeek V3.1 is an August 2025 hybrid model that combines the capabilities of V3 and R1, supporting both thinking and non-thinking modes via chat template switching. It features 671B parameters (37B activated), 128K context, and significantly improved tool-calling and agent capabilities.

by DeepSeek | $0.21/M input | $0.79/M output

Imagen 4 Fast is Google's speed-optimized text-to-image model offering generation up to 10x faster than Imagen 3 at just $0.02 per image. It's ideal for rapid prototyping, high-volume tasks, and iterative exploration while maintaining improved text rendering and style versatility.

by Google | $0.02/image

Imagen 4 Ultra is Google's highest-fidelity text-to-image model designed for professional-grade realism with superior prompt adherence and nuanced interpretation of complex scenes. It delivers exceptional detail in textures, lighting, and atmosphere with 2K resolution output at $0.06 per image.

by Google | $0.0602/image

Imagen 4 is Google DeepMind's flagship text-to-image generation model, available through the Gemini API and Google AI Studio. It delivers significant improvements over Imagen 3, particularly in rendering text, typography, and fine details like intricate fabrics and textures. The model supports output up to 2K resolution across a range of aspect ratios, generating images in roughly 2.5 seconds. A Fast variant optimized for high-volume use runs at $0.02 per image, while the standard model is $0.04 and the Ultra tier—built for precise prompt adherence—is $0.06. In human evaluations on GenAI-Bench, Imagen 4 scored highly against other leading image generation models on overall preference. All outputs are embedded with Google's SynthID watermark for AI-content traceability. It's a strong fit for developers building creative tools, marketing asset pipelines, or any application requiring reliable, high-quality image generation from text prompts.

by Google | $0.04/image

Mistral Medium 3.1 (August 2025) is a frontier-class multimodal model with improved tone and performance. It features 128K context, native vision, and enhanced reasoning for STEM and enterprise workflows at competitive pricing.

by Mistral AI | $0.4/M input | $2/M output

Mistral Medium 3.1 is Mistral's frontier-class multimodal model released August 2025 with 128K context. It delivers near-frontier performance at $0.4/$2 per million tokens, excelling in reasoning, coding, and enterprise workflows.

by Mistral AI | $0.4/M input | $2/M output

Jamba Large 1.7 is AI21 Labs' flagship open-weight language model, built on a hybrid SSM-Transformer (Mamba-Transformer) architecture with a Mixture of Experts design — 398B total parameters with 94B active during inference. Its standout feature is a 256K-token context window, making it well suited for processing lengthy documents, contracts, and knowledge bases. The model supports function calling, JSON mode, and nine languages including English, Spanish, French, German, and Arabic. Jamba Large 1.7 emphasizes grounding and instruction-following, delivering contextually faithful responses with strong steerability. It generates output at roughly 69 tokens per second via the AI21 API. It targets enterprise workflows in domains like finance, healthcare, and legal — where long-context accuracy and data control matter most.

by AI21 Labs | $2/M input | $8/M output

GPT-5 is OpenAI's unified reasoning system combining a fast model, a deeper thinking model, and an automatic router. It achieves 45% fewer factual errors than GPT-4o and sets state-of-the-art scores on math, coding, and health benchmarks.

by OpenAI | $1.25/M input | $10/M output

GPT-5 Chat is the non-reasoning version of GPT-5 used in ChatGPT, designed for conversational interactions. It's available as gpt-5-chat-latest in the API and provides fast responses without extended thinking.

by OpenAI | $1.25/M input | $10/M output

GPT-5 Mini is a faster, more cost-efficient version of GPT-5 optimized for well-defined tasks and precise prompts. It provides a balance between performance and speed for everyday use cases.

by OpenAI | $0.25/M input | $2/M output

GPT-5 Nano is OpenAI's fastest and cheapest GPT-5 variant, ideal for summarization and classification tasks. It offers extremely low latency for high-volume, simple inference workloads.

by OpenAI | $0.05/M input | $0.4/M output

GPT-5 Codex is a version of GPT-5 optimized for agentic coding tasks in Codex or similar environments. It's designed for software engineering workflows with enhanced code generation capabilities.

by OpenAI | $1.25/M input | $10/M output

GPT-OSS 120B is OpenAI's most powerful open-weight model under Apache 2.0 license, achieving near-parity with o4-mini on reasoning benchmarks. It has 117B total parameters with 5.1B active, fitting on a single H100 GPU.

by OpenAI | $0.04/M input | $0.17/M output

GPT-OSS 20B is OpenAI's smaller open-weight model for lower latency and local inference, matching o3-mini on common benchmarks. It requires only 16GB of memory and runs on consumer hardware.

by OpenAI | $0.03/M input | $0.14/M output

GPT-OSS 120B is an open-weight reasoning model built by OpenAI and released under the Apache 2.0 license, available here through BytePlus's hosting. It is built for agentic workflows, including tool use, web search, and code execution, with adjustable reasoning effort and full chain-of-thought output. It supports structured outputs and is compatible with OpenAI's Responses API format. On reported benchmarks, it scores 90% on MMLU and 96.6% on AIME 2024, and 80-81% on GPQA Diamond, putting it near parity with OpenAI's o4-mini on core reasoning tasks. It also outperforms o3-mini on competition coding, tool calling (TauBench), and general problem solving, and scores above o4-mini on HealthBench. Because the weights are open, it suits teams wanting OpenAI-level reasoning quality without the licensing restrictions of closed models like GPT-4.

by BytePlus | $0.1/M input | $0.5/M output

Qwen Image is a 20B-parameter image generation foundation model from Alibaba's Qwen series, built for text-to-image generation, image editing, and image understanding tasks. Its standout capability is high-fidelity text rendering — it accurately places readable text in both English and Chinese within generated images, making it especially strong for posters, slides, and design-heavy visuals. Beyond text, it supports a wide range of styles from photorealism to anime, and handles advanced editing operations like style transfer, object insertion/removal, and in-image text modification. The model also performs image understanding tasks including object detection, segmentation, depth estimation, and super-resolution. A versatile choice for developers who need generation, editing, and visual analysis in a single model. Licensed under Apache 2.0.

by Qwen | $0.0058/image

MiniMax Hailuo 02 is a next-generation AI video model ranked #2 globally, featuring native 1080p output and advanced physics simulation for realistic motion including gravity, fluid dynamics, and complex movements like gymnastics. It uses Noise-aware Compute Redistribution (NCR) architecture for 2.5x improved efficiency, with 3x more parameters and 4x more training data than its predecessor. The model supports both text-to-video and image-to-video generation with clips up to 10 seconds.

by MiniMax

Qwen3 30B A3B Instruct (2507) is the July 2025 updated instruction-tuned version with improved capabilities in reasoning, coding, and tool usage at high efficiency.

by Qwen | $0.1/M input | $0.3/M output

GLM-4.5 is Zhipu AI's flagship 355B-parameter open-source model (32B active) designed for agentic AI applications with dual thinking/non-thinking modes. It excels at reasoning, coding, and tool use, ranking 3rd globally among all models on combined benchmarks under MIT license.

by Z.AI | $0.6/M input | $2.2/M output

GLM-4.5-Air is a compact 106B-parameter variant (12B active) of GLM-4.5, offering competitive agentic performance with significantly lower resource requirements. It supports the same dual reasoning modes and 128K context window as its larger sibling.

by Z.AI | $0.2/M input | $1.1/M output

GLM-4.5-AirX is the ultra-fast inference variant of Z.ai's GLM-4.5-Air, a 106B-parameter Mixture-of-Experts model with 12B active parameters per forward pass. It shares the same architecture and 128K-token context window as GLM-4.5-Air but is optimized for maximum throughput and minimal latency. GLM-4.5-Air itself delivers strong results — scoring 59.8 across 12 industry benchmarks and outperforming models like Gemini 2.5 Flash and Qwen3-235B on reasoning evaluations. AirX preserves that capability while targeting low-latency, high-concurrency production scenarios. Best suited for real-time agent pipelines, high-volume chat, and latency-sensitive coding assistance where the full GLM-4.5's throughput is insufficient but you still need competitive reasoning and tool-use performance.

by Z.AI | $1.1/M input | $4.5/M output

GLM-4.5-Flash is the free tier in Z.ai's GLM-4.5 model family, optimized for coding, reasoning, and agent tasks. It shares the hybrid reasoning architecture of the broader GLM-4.5 series, supporting both a thinking mode for complex multi-step problems and a non-thinking mode for instant responses. With a 128K-token context window and native support for function calling, structured output, and streaming, it provides a capable baseline for developers prototyping agent workflows or building cost-sensitive applications. It integrates with coding agent frameworks like Claude Code and Roo Code. An excellent starting point for teams evaluating the GLM-4.5 ecosystem — no cost to experiment, with a clear upgrade path to GLM-4.5 or GLM-4.5-X for heavier workloads.

by Z.AI | $0.1/M input | $0.1/M output

GLM-4.5-X is the high-performance, ultra-fast inference variant of Z.ai's flagship GLM-4.5 model. It retains the full 355B-parameter MoE architecture (32B active) and 128K-token context window while being tuned for significantly faster response times — exceeding 100 tokens per second in real-world tests. GLM-4.5 itself ranks among the top models globally across 12 benchmarks spanning reasoning, coding, and agentic tasks, with an aggregate score of 63.2. The X variant delivers that same capability ceiling with latency suitable for interactive applications. Designed for production workloads where both quality and speed matter — real-time coding agents, interactive tool-use pipelines, and high-concurrency deployments that can't afford the response time of the standard GLM-4.5 endpoint.

by Z.AI | $2.2/M input | $8.9/M output

GLM-4.5X is the high-speed serving of GLM-4.5, the flagship agentic model from Z.ai (Zhipu AI). It keeps the same 128K-token context window and 96K maximum output as the standard endpoint but is served for low latency, with generation speed exceeding 100 tokens per second in real-world tests. The underlying GLM-4.5 ranks near the top globally on an aggregate of 12 benchmark suites (including MMLU Pro, AIME24, MATH 500, and GPQA) and is optimized for tool invocation, web browsing, software engineering, and front-end development. It supports function calling and both thinking and non-thinking response modes. Choose it over standard GLM-4.5 when latency matters more than cost, for interactive coding assistants, real-time agent loops, and high-concurrency chat. Pricing is $2.20 per million input tokens and $8.90 per million output, roughly four times the standard GLM-4.5 rate.

by Z.AI | $2.2/M input | $8.9/M output

Qwen3 Coder Flash is a cost-effective coding model balancing performance and speed, suitable for scenarios requiring fast responses at lower cost while maintaining coding quality.

by Qwen | $0.3/M input | $1.5/M output

Qwen Flash is Alibaba's latency-optimized general-purpose language model, designed as the successor to Qwen Turbo for cost-efficient, high-throughput workloads. It offers a 1 million token context window with native support for context caching, making repeated or large-context requests significantly cheaper. The model supports function calling and is accessible via an OpenAI-compatible API through Alibaba Cloud Model Studio. Qwen Flash is a strong choice for developers running high-volume production tasks — classification, extraction, summarization, and lightweight agentic pipelines — where low latency and predictable pricing matter more than peak reasoning capability. Its flexible tiered pricing and context cache support make it especially cost-effective at scale.

by Qwen | $0.05/M input | $0.4/M output

Qwen3 235B A22B Thinking (2507) is the reasoning-enhanced variant using extended chain-of-thought processing for complex math, coding, and logical problems with enhanced performance.

by Qwen | $0.23/M input | $2.3/M output

Qwen3 Coder Plus is the strongest Qwen coding API model, ideal for complex project generation and in-depth code reviews with up to 1M token context support.

by Qwen | $1/M input | $5/M output

UI-TARS 1.5 7B is a multimodal vision-language agent by ByteDance optimized for GUI automation across desktop, web, mobile, and game environments. It uses reinforcement learning-based reasoning to plan and execute actions on graphical interfaces. The model achieves state-of-the-art results on benchmarks like OSWorld, WebVoyager, and AndroidWorld.

by ByteDance | $0.1/M input | $0.2/M output

Qwen3 235B A22B (2507) is the July 2025 updated version with significant improvements in instruction following, reasoning, coding, tool usage, and 256K long-context understanding.

by Qwen | $0.09/M input | $0.55/M output

Voxtral Small is a 24B parameter speech understanding model built on Mistral Small 3.1 under Apache 2.0. It supports 30-minute transcription, 40-minute audio understanding, Q&A, summarization, and function calling from voice in 8+ languages.

by Mistral AI | $0.1/M input | $0.3/M output

Seed 1.6 Flash is ByteDance's faster, lower-cost tier of the Seed 1.6 family, accessible here through BytePlus's Model Ark hosting. It shares the family's 256,000-token context window and reads text, image, and video input, and it launched alongside Seed 1.6 and Seed 1.6-Thinking as part of ByteDance's Doubao 1.6 rollout, which cut token pricing sharply from the prior generation. ByteDance built its Seed-1.6-Embedding model on top of Seed 1.6 Flash, reusing its multimodal understanding of text, images, and video as the base for that embedding model. Tool calling is supported, and input and output pricing run well below the full Seed 1.6 model, so it targets high-volume, latency-sensitive use cases such as real-time chat, agent loops with frequent tool calls, and media classification. Teams that need the strongest reasoning should use Seed 1.6 or Seed 1.6-Thinking instead.

by BytePlus | $0.08/M input | $0.3/M output

Kimi K2 is a trillion-parameter Mixture-of-Experts model by Moonshot AI, activating 32 billion parameters per token. Designed as a non-thinking model optimized for agentic capabilities, it excels at tool use, code generation, and autonomous problem-solving with a 128K token context window. On benchmarks, K2 scored 65.8% on SWE-bench Verified, 75.1% on GPQA-Diamond, 49.5% on AIME 2025, and 66.1 on Tau2-bench — surpassing most open- and closed-source models in non-thinking settings. It ranked as the #1 open-source model on the LMSYS Arena leaderboard upon release in July 2025. K2 is well suited for developers building AI agents and tool-calling pipelines who need strong coding and reasoning without extended thinking overhead.

by Moonshot AI | $0.57/M input | $2.3/M output

Kimi K2 Instruct is the instruction-tuned version of Moonshot AI's Kimi K2, released in July 2025 as a general-purpose chat and agentic model. Moonshot describes it as reflex-grade, meaning it answers directly without extended thinking, which keeps latency low in tool-calling loops. It scores 65.8% on SWE-bench Verified with agentic tools (single attempt), 70.6% on Tau2-bench retail, and 76.5% on AceBench, close to Claude 4 on tool-use benchmarks at the time of release. Tool calling is built in; you pass a list of available tools and the model decides when and how to invoke them. This route offers a 128K context window. It suits developers building agents, coding assistants, and multi-step tool pipelines that need autonomous execution without chain-of-thought token overhead.

by Moonshot AI | $0.7/M input | $2.5/M output

Grok 4 is xAI's flagship reasoning model released July 2025, trained with unprecedented reinforcement learning scale on 200,000 GPUs. It features native tool use, real-time search integration, and Grok 4 Heavy achieves 50% on Humanity's Last Exam benchmark.

by xAI | $3/M input | $15/M output

Grok 4 0709 is the July 9, 2025 snapshot of xAI's flagship reasoning model, trained with reinforcement learning to use tools like a code interpreter and web browsing. It features a 256K context window, native tool use, parallel tool calling, and support for both image and text inputs.

by xAI | $3/M input | $15/M output

Dolphin Mistral 24B Venice Edition is a 24 billion parameter, uncensored general-purpose language model fine-tuned from Mistral Small 24B (Instruct-2501), built by Cognitive Computations (the Dolphin project, founded by Eric Hartford) in collaboration with Venice.ai. This paid API variant offers a 128K context window and an April 2024 knowledge cutoff. On Venice's uncensored benchmark, it recorded a 2.2% refusal rate, the lowest among tested models and well below Llama 4 Maverick (11.11%), Gemini 2.5 Flash Preview (53.33%), GPT-4o-mini (64.44%), and Claude Sonnet 3.7 (71.11%). While base Mistral Small 24B leaned STEM-heavy, this fine-tune adds strong creative writing and storytelling, with consistent character and narrative memory, plus steerable tone control that defaults to neutral and polite. Suited for developers who need minimal refusals, custom alignment, or unrestricted content generation via API.

Hunyuan A13B Instruct is an open-source large language model from Tencent built on a fine-grained Mixture-of-Experts (MoE) architecture, with 80B total parameters and 13B active during inference. It natively supports a 256K-token context window. It performs competitively with OpenAI o1 and DeepSeek R1 across math, science, and reasoning benchmarks, scoring 87.3 on AIME 2024, 89.1 on BBH, and 84.7 on ZebraLogic. Hunyuan A13B particularly excels at agentic tasks and tool use, leading on benchmarks like BFCL-v3 (78.3) and ComplexFuncBench (61.2). It's a strong choice for developers building agent workflows, long-context applications, or cost-sensitive reasoning pipelines.

by Tencent | $0.14/M input | $0.57/M output

Morph V3 Fast is a specialized code-editing model built by Morph, designed to serve as the execution layer in AI-assisted development workflows. Rather than generating code from scratch, it applies edits suggested by frontier reasoning models like Claude or GPT-4o to existing code files. It processes at approximately 10,500 tokens per second with around 96% accuracy on code transformations, making it one of the fastest options for automated code apply tasks. The model supports an 81,920-token context window and up to 38,000 output tokens. Morph V3 Fast is built for high-volume, latency-sensitive pipelines where code edits need to be applied rapidly and cheaply.

by Morph | $0.8/M input | $1.2/M output

Morph V3 Large is Morph's high-accuracy code apply model, optimized for complex and precise code transformations. Like its faster sibling, it acts as the execution layer in agentic coding workflows — taking edit suggestions from reasoning models and merging them into existing code. It achieves approximately 98% accuracy on code transformations at speeds around 4,500 tokens per second. The model supports a 262,144-token context window with up to 131,100 output tokens, allowing it to process entire codebases or large files in a single request. Morph V3 Large is the better choice when edit correctness matters more than raw speed — particularly for production codebases or complex multi-file changes where a 2% accuracy gap can mean hundreds of broken edits at scale.

by Morph | $0.9/M input | $1.9/M output

Qwen-MT Lite is the lowest-cost tier of Alibaba's Qwen-MT machine translation family, built for simple, latency-sensitive translation such as real-time chat and live comment translation. It supports 31 languages, fewer than the 92 covered by Qwen-MT Plus and Qwen-MT Turbo, and Alibaba's model comparison rates it as basic quality at the fastest speed and lowest cost of the tiers. Unlike the higher tiers, it does not support domain prompting. It does support incremental streaming output, where each response contains only newly generated content. Best suited for developers translating short, high-volume text (chat messages, comments, UI strings) where response speed and price matter more than translation fidelity. For terminology control or higher output quality, Qwen-MT Plus or Turbo are the better fit.

by Qwen | $0.12/M input | $0.36/M output

ERNIE 4.5 VL 424B A47B is Baidu's largest multimodal vision-language model with 424B total parameters and 47B active per token. It supports up to 131K context tokens and excels at visual reasoning, document/chart understanding, and visual question answering with both thinking and non-thinking modes. In thinking mode, it approaches or surpasses OpenAI o1 on reasoning benchmarks like MathVista, MMMU, and VisualPuzzle.

by Baidu | $0.42/M input | $1.25/M output

Anubis 70B v1.1 is a finetune of Meta's Llama 3.3 70B Instruct by TheDrummer, an independent model creator known for roleplay and creative-writing finetunes. The finetune targets roleplay and character-driven writing. TheDrummer describes v1.1 as a substantial overhaul of v1.0, with improved character adherence (characters keep their personality and voice across a conversation) and reduced alignment compared to the base model. The model uses the standard Llama 3 chat template. It suits developers building roleplay or interactive-fiction applications who want a 70B-class model tuned for character consistency rather than general assistant tasks.

Seedance 1.0 Lite is ByteDance's speed-optimized AI video generation model that creates 5 second videos at 480p-720p resolution from text prompts or images. It supports text-to-video and image-to-video generation with smooth motion and multi-shot narrative capabilities, designed for fast iteration and experimentation.

by ByteDance

Seedance 1.0 Pro is ByteDance's professional-grade AI video generation model that produces cinematic 1080p videos from text or images. It excels at multi-shot storytelling with consistent subjects and visual style across scenes, featuring smooth motion, rich details, and advanced prompt following for production-quality content.

by ByteDance

Mistral Small 3.2 improves on 3.1 with better instruction following (84.78% vs 82.75%), reduced infinite generations (1.29% vs 2.11%), and more robust function calling. It maintains the 24B/128K context architecture under Apache 2.0.

by Mistral AI | $0.09/M input | $0.25/M output

MiniMax-M1 is the world's first open-source hybrid-attention reasoning model, featuring a 1 million token context window and 80K reasoning output budget. It excels in software engineering, long-context tasks, and complex reasoning while being trained with an efficient CISPO reinforcement learning algorithm.

by MiniMax | $0.55/M input | $2.2/M output

Gemini 2.5 Flash-Lite is Google's cost-optimized version of 2.5 Flash, designed for high-volume tasks like classification, translation, and intelligent routing. It delivers efficient performance for cost-sensitive, high-scale operations.

by Google | $0.1/M input | $0.4/M output

Gemini 2.5 Flash-Lite Preview 06-17 is the June 17, 2025 preview snapshot of Gemini 2.5 Flash-Lite, the fastest and lowest-priced model in Google's Gemini 2.5 family. It accepts text, image, audio, and video input, has a 1M token context window, and supports tool calling. Thinking is off by default and can be enabled with a thinking budget for harder tasks. Google reported lower latency than Gemini 2.0 Flash and 2.0 Flash-Lite, and higher quality than 2.0 Flash-Lite on coding, math, science, reasoning, and multimodal benchmarks. Artificial Analysis measured it at about 5x cheaper and 1.7x faster than Gemini 2.5 Flash, with lower intelligence. It fits high-volume, latency-sensitive work such as translation and classification. The stable gemini-2.5-flash-lite superseded this preview at general availability on July 22, 2025, so use this ID only if you need this exact snapshot.

by Google | $0.1/M input | $0.4/M output

Skylark Pro SC 250615 is a June 2025 snapshot of Skylark Pro, ByteDance's flagship chat model line offered through its BytePlus ModelArk platform. Skylark was ByteDance's original large language model family, made commercially available in 2023, and is the international counterpart to the Doubao models. BytePlus positions Skylark Pro for complex question answering, summarization, creative writing, and role-play, and publishes a role creation guide for building persona-driven and multi-character chat. Role-play platforms such as Moescape have adopted it as a character chat model. The model has a 128,000-token context window, and this deployment supports up to 12,000 output tokens per response. BytePlus documentation lists function calling and structured output among the Skylark Pro line's capabilities, with pricing of $0.40 per million input tokens and $1.60 per million output tokens.

by ByteDance | $0.4/M input | $1.6/M output

FLUX.1 Kontext Max is the highest-quality model in the Kontext series, optimized for iteratively modifying existing images via text prompts with maximum fidelity. It offers the best editing consistency and prompt following among Kontext variants.

by Black Forest Labs | $0.08/image

FLUX.1 Kontext Pro is a production-grade in-context image generation and editing model that balances quality and speed. It powers integrations in Adobe Photoshop's Generative Fill and Meta's platforms.

by Black Forest Labs | $0.04/image

OpenAI o3 Pro is a version of o3 designed to think longer and provide the most reliable responses for challenging questions. It's recommended when reliability matters more than speed.

by OpenAI | $20/M input | $80/M output

Gemini 2.5 Pro Preview is the preview version of Google's most advanced reasoning model with state-of-the-art coding and complex task performance. It features Deep Think mode, 1M token context, and advanced multimodal capabilities.

by Google | $1.25/M input | $10/M output

Gemini 2.5 Pro Preview 06-05 is the June 5, 2025 snapshot of Google's Gemini 2.5 Pro reasoning model, and the final preview before general availability. The stable gemini-2.5-pro release contains no changes from this snapshot, so the two behave identically. At release it scored 1470 Elo on LMArena (up 24 points over the prior 05-06 preview) and 1443 on WebDevArena (up 35 points), led the Aider Polyglot coding benchmark, and posted top-tier results on GPQA and Humanity's Last Exam. Google also improved the style and structure of its responses based on developer feedback on earlier previews. It supports a 1M token context window, tool calling, and text, image, audio, and video input. As a dated snapshot its behavior is fixed; new work can use the stable gemini-2.5-pro ID instead.

by Google | $1.25/M input | $10/M output

DeepSeek R1-0528 is the May 2025 major update to R1, featuring dramatically improved reasoning depth with nearly double the thinking tokens (23K vs 12K average) and approaching performance of O3 and Gemini 2.5 Pro. It adds function calling support, reduced hallucinations, and improved AIME accuracy from 70% to 87.5%.

by DeepSeek | $0.5/M input | $2.15/M output

Seedance 1.0 Pro is ByteDance's original professional-tier Seedance video model, released in June 2025 and available here through BytePlus's hosting. It generates text-to-video and image-to-video clips with native multi-shot generation, keeping subjects and visual style consistent across cuts rather than treating each shot independently. At launch it topped the Artificial Analysis community leaderboard for both text-to-video and image-to-video, ahead of Google's Veo 3, OpenAI's Sora, and Kling 2.0. ByteDance's technical report also cites roughly a 10x inference speedup, generating a 5-second 1080p clip in about 41 seconds. Through this API it supports 480p, 720p, and 1080p at up to 12 seconds and 24fps, and suits production work such as marketing and narrative content that needs multi-shot consistency.

by BytePlus

Claude Opus 4 is the flagship model from the May 2025 Claude 4 launch, designed for complex long-running tasks. It can work continuously for several hours (7+ hour coding sessions demonstrated) and leads on coding benchmarks at 72.5% SWE-bench.

by Anthropic | $15/M input | $75/M output

Claude Sonnet 4 is the May 2025 successor to Sonnet 3.7 with enhanced steerability and coding (72.7% SWE-bench). It excels at following complex instructions precisely and autonomous multi-feature app development with near-zero navigation errors.

by Anthropic | $3/M input | $15/M output

Gemma 3n E4B Instruct is Google's mobile-optimized model with a 4B active memory footprint containing a nested 2B submodel for flexible quality-latency tradeoffs. It supports real-time multimodal processing on edge devices.

by Google | $0.06/M input | $0.12/M output

Imagen 4 Preview is the preview version of Google's flagship text-to-image diffusion model featuring photorealistic detail, improved typography, and support for up to 2K resolution. It balances quality and cost at $0.04 per image, making it suitable for a wide variety of creative tasks.

by Google | $0.04/image

Google Veo 3 is Google DeepMind's advanced AI video model that generates high-quality videos with native synchronized audio including dialogue, sound effects, and ambient noise directly from text prompts. It delivers state-of-the-art results in physics, realism, and prompt adherence with cinematic quality 8-second clips at up to 1080p resolution.

by Google

Google Veo 3 with Audio is the audio-enabled configuration of Veo 3 that generates synchronized sound effects, dialogue, ambient noise, and music natively alongside video content. It produces complete audiovisual experiences from text prompts, eliminating the need for separate audio post-production.

by Google

Google Veo 3 Fast is a speed-optimized variant of Veo 3 that generates videos approximately 2x faster at 60-80% lower cost while maintaining high visual quality. It's designed for rapid iteration, prototyping, and cost-efficient production workflows at 720p resolution.

by Google

Google Veo 3 Fast with Audio is the audio-enabled version of the speed-optimized Veo 3 Fast model, combining faster generation times and lower costs with native synchronized audio generation. It delivers sound effects, dialogue, and ambient audio while optimizing for speed and affordability in production workflows.

by Google

Gemini 2.5 Flash Preview 05-20 is the May 20, 2025 snapshot of Google's Gemini 2.5 Flash hybrid reasoning model, announced at Google I/O as the replacement for the earlier 04-17 preview. Google reported improvements across reasoning, multimodality, code, and long context, along with 20-30% lower token use in its evaluations. On Artificial Analysis's intelligence index it scores 65 with thinking on (up from 60) and 53 with thinking off (up from 49), putting its non-reasoning mode roughly level with GPT-4.1 and DeepSeek V3 0324. Thinking can be turned on or off per request, with a thinking budget of up to 24,576 tokens to trade off cost, latency, and quality. It accepts text, image, audio, and video input with a 1M token context window, supports function calling, and became the stable Gemini 2.5 Flash release unchanged in June 2025.

by Google | $0.3/M input | $2.5/M output

FLUX.2 Dev is a 32B parameter open-weight flow matching transformer for text-to-image generation and multi-reference image editing. It uses Mistral-3 24B as its vision-language backbone and features a new VAE, improved typography, and support for up to 10 reference images.

by Black Forest Labs | $0.03/image

FLUX.2 Flex is a specialized FLUX.2 variant focused on typography accuracy and fine detail preservation, with a configurable 'steps' parameter for trading off between quality and speed. It exposes advanced controls like guidance and step count for precise diffusion tuning.

by Black Forest Labs | $0.03/image

FLUX.2 Max is the most capable model in the FLUX.2 family, delivering the highest editing consistency, strongest prompt following, and best photorealism. It uniquely supports grounded generation with real-time web search to visualize current events and trending content.

by Black Forest Labs | $0.07/image

FLUX.2 Pro is the production-grade FLUX.2 model balancing high quality and affordability, designed for professional image generation and editing workflows. It uses fixed optimal inference parameters for consistent output without manual tuning.

by Black Forest Labs | $0.03/image

Vidu Q1 is a high-performance generative video model from ShengShu Technology that produces cinematic 1080p videos up to 5 seconds with integrated AI-generated audio. It features a First-to-Last Frame system for seamless transitions between unrelated images and supports up to 7 reference images for multi-character consistency. The model excels at anime-style content and includes built-in 48kHz sound effects and background music generation from text prompts.

by Vidu

Gemini 2.5 Pro Preview (May 6) is a dated preview snapshot of Google's flagship reasoning model with improvements in code and function calling. It offers advanced reasoning capabilities for complex enterprise use cases.

by Google | $1.25/M input | $10/M output

Arcee Virtuoso Large is a 72-billion-parameter general-purpose language model from Arcee AI, built on Qwen2.5-72B and post-trained using DeepSeek R1 distillation, multi-epoch supervised fine-tuning, and DPO/RLHF alignment. It is designed for cross-domain reasoning, enterprise question answering, creative writing, and long-document comprehension, with a 128k context window that enables processing entire codebases or lengthy documents in a single API call. Virtuoso Large is Arcee's flagship dense general-purpose model — a solid default choice for developers who need reliable, broad-capability performance without the routing complexity of MoE architectures.

by Arcee AI | $0.75/M input | $1.2/M output

Qwen3 0.6B is the smallest dense model in Alibaba's Qwen3 family, released in April 2025 with 0.6 billion parameters (0.44 billion non-embedding). Like the larger Qwen3 models, it supports hybrid thinking modes, switching between step-by-step reasoning and a faster direct-answer mode per request, including with /think and /no_think tags in prompts. The Qwen3 technical report lists 55.6 on MMLU-Redux and 59.6 on GSM8K in thinking mode, ahead of Qwen2.5-0.5B-Instruct (41.6 on GSM8K). The family was pretrained on 36 trillion tokens covering 119 languages, and the model supports tool calling, with the Qwen-Agent framework recommended for agent workloads. At this size it fits high-volume, cost-sensitive tasks such as classification, routing, and drafting, and it is also used as a draft model for speculative decoding with larger Qwen3 models. Complex multi-step math stays well below what 1B-plus models manage.

by Qwen | $0.11/M input | $0.42/M output

Wan 2.2 I2V A14B is an open-source image-to-video generation model that transforms static images into 5-second videos at 480P or 720P resolution. It uses a Mixture-of-Experts (MoE) architecture with dual 14B-parameter experts to achieve stable video synthesis with reduced unrealistic camera movements and enhanced support for diverse stylized scenes.

by Wan AI

Wan 2.2 T2V A14B is an open-source text-to-video generation model that creates 5-second videos at 480P or 720P resolution from text prompts. Built with a Mixture-of-Experts (MoE) architecture featuring specialized high-noise and low-noise experts, it delivers cinematic-quality output with granular control over lighting, composition, and motion.

by Wan AI

GPT Image 1 is OpenAI's natively multimodal image generation model released in April 2025, built on GPT-4o architecture to accept both text and image inputs. It excels at text rendering, detailed instruction following, and photorealistic output with support for image editing and inpainting. The model uses an autoregressive approach rather than diffusion, representing a significant advancement over the DALL·E series.

by OpenAI | $0.011/image

PixVerse V5 is an AI video generation model that converts text or images into cinematic-quality videos with smooth motion, fast rendering speeds, and enhanced prompt adherence. It excels at creating high-fidelity videos with natural camera movements and consistent visual styling across frames. The model is ranked among the top performers in image-to-video and text-to-video benchmarks.

by PixVerse

Gemini 2.5 Flash Image (codenamed Nano Banana) is Google's state-of-the-art multimodal model for fast, conversational image generation and editing with low latency. It maintains character consistency across prompts, enables precise local edits via natural language, and supports multi-image composition and fusion.

by Google | $0.039/image

Gemini 2.5 Flash Preview is the April 2025 preview release (04-17) of Google's first fully hybrid reasoning model. Developers can turn thinking on or off per request and set a thinking budget in tokens to trade off quality, cost, and latency. With the budget at zero it runs at cost and latency comparable to Gemini 2.0 Flash while scoring higher on benchmarks. In Google's reported results it scored 12.1% on Humanity's Last Exam, ahead of Claude 3.7 Sonnet (8.9%) and DeepSeek R1 (8.6%) but behind o4-mini (14.3%), plus 78.3% on GPQA Diamond and 78.0% on AIME 2025. It accepts text, image, audio, and video input with a 1M token context window. It suits high-volume work like chat, summarization, and data extraction where reasoning depth needs to be controlled. As a preview endpoint it has since been superseded by the stable Gemini 2.5 Flash release.

by Google | $0.3/M input | $2.5/M output

OpenAI o3 is a powerful reasoning model that pushes the frontier in coding, math, science, and visual perception. It can agentically use all ChatGPT tools and makes 20% fewer major errors than o1 on difficult tasks.

by OpenAI | $2/M input | $8/M output

OpenAI o4 Mini is a fast, cost-efficient reasoning model optimized for coding and visual tasks. It achieves remarkable performance for its size with full tool access in ChatGPT, succeeded by GPT-5 Mini.

by OpenAI | $1.21/M input | $4.84/M output

OpenAI o4 Mini High is a higher-intelligence version of o4-mini available in the ChatGPT model picker. It provides enhanced reasoning at the cost of longer response times.

by OpenAI | $1.1/M input | $4.4/M output

Kling 2.1 Master is Kuaishou's premium AI video generation model featuring 1080p output, advanced 3D spatiotemporal attention for cinematic-grade realism, and superior prompt adherence. It supports both text-to-video and image-to-video with refined facial modeling and complex motion dynamics ideal for professional filmmakers and advertisers.

by Kling

GLM-4-32B-0414-128K is a 32B-parameter dense language model from Z.ai with an extended 128K-token context window. Pre-trained on 15 trillion tokens of high-quality data — including substantial reasoning-focused synthetic data — it was further refined with rejection sampling and reinforcement learning for instruction following, code generation, and function calling. It supports bilingual Chinese-English usage and is optimized for tasks like tool use, search-grounded Q&A, and structured output generation. Performance is competitive with models in the GPT and DeepSeek V3/R1 class at a fraction of the parameter count. A strong choice for cost-sensitive workloads that need long-context reasoning, multi-file code editing, or reliable JSON output without stepping up to the larger MoE models in the GLM family.

by Z.AI | $0.1/M input | $0.1/M output

GPT-4.1 is OpenAI's smartest non-reasoning model, excelling at instruction following and tool calling with a 1M token context window. It outperforms GPT-4o across coding and multimodal tasks with a June 2024 knowledge cutoff.

by OpenAI | $2/M input | $8/M output

GPT-4.1 Mini is a smaller, faster version of GPT-4.1 that matches or exceeds GPT-4o performance while reducing latency by nearly half and cost by 83%. It features a 1M token context window and strong coding capabilities.

by OpenAI | $0.4/M input | $1.6/M output

GPT-4.1 Nano is OpenAI's fastest and cheapest model, designed for low-latency tasks like classification and autocompletion. It features a 1M token context window and scores 80.1% on MMLU despite its small size.

by OpenAI | $0.1/M input | $0.4/M output

HiDream I1 Dev is a guidance-distilled, 17-billion-parameter text-to-image model from HiDream AI, built on a sparse Diffusion Transformer architecture with dynamic Mixture-of-Experts layers. It runs in approximately 28 diffusion steps, placing it between the Full and Fast variants in the speed-quality tradeoff. Because it is distillation-trained, negative prompts are not required — the classifier-free guidance scale should be set to 1.0 during sampling. I1 Dev is well-suited for iterative development workflows, concept exploration, and production pipelines where response time matters but image fidelity cannot be fully sacrificed. It also supports LoRAs for style control.

by HiDream | $0.0045/image

HiDream I1 Fast is the lowest-latency variant of HiDream AI's 17-billion-parameter text-to-image model family, completing generation in as few as 14-16 diffusion steps. Like the Dev variant, it is distillation-trained and does not require negative prompts. It is the best-fit variant for latency-sensitive API integrations, delivering strong image quality for its step count. I1 Fast is the right choice for real-time generation features, high-throughput batch pipelines, or user-facing products where responsiveness is the primary constraint. Maximum fidelity is reserved for the Full variant.

by HiDream | $0.0032/image

HiDream I1 Full is the flagship text-to-image model from HiDream AI, a 17-billion-parameter sparse Diffusion Transformer that delivers the highest output quality in the I1 family through 50+ diffusion steps. On verified benchmarks, it scores 0.83 on GenEval (vs. 0.67 for DALL-E 3 and 0.66 for FLUX.1-dev), 85.89 on DPG-Bench for prompt adherence, and 33.82 on HPS v2.1 (vs. 32.47 for FLUX.1-dev and 31.44 for DALL-E 3). It supports negative prompts with a guidance scale of 5.0 for finer output control. I1 Full is the best choice when image quality, prompt fidelity, and detail richness are the top priorities — ideal for asset generation, creative production, and high-stakes visual content pipelines.

by HiDream | $0.009/image

Llama 4 Maverick is Meta's 400 billion total parameter MoE model with 17B active parameters and 128 experts, supporting 1M token context. It's natively multimodal with state-of-the-art performance on coding, reasoning, and image understanding tasks.

by Meta Llama | $0.15/M input | $0.6/M output

Llama 4 Scout is Meta's efficient 109 billion parameter MoE model with 17B active parameters and 16 experts, featuring an industry-leading 10M token context window. It fits on a single H100 GPU and handles multimodal text and image inputs.

by Meta Llama | $0.1/M input | $0.3/M output

Qwen3 14B is a dense language model with hybrid thinking/non-thinking modes, matching Qwen2.5-32B performance. It supports 119 languages and excels in math, coding, and reasoning tasks.

by Qwen | $0.35/M input | $1.4/M output

Qwen3 235B A22B is the flagship MoE model with 235B total and 22B active parameters, rivaling DeepSeek-R1 and o1. It features hybrid thinking modes and supports 119 languages with strong agentic capabilities.

by Qwen | $0.7/M input | $2.8/M output

Qwen3 32B is a dense language model matching Qwen2.5-72B performance with hybrid thinking/non-thinking modes. It excels in STEM, coding, and reasoning while supporting 119 languages.

by Qwen | $0.7/M input | $2.8/M output

Qwen3 8B is a dense model matching Qwen2.5-14B performance with hybrid thinking modes and 128K context. It offers strong reasoning, coding, and multilingual capabilities in a mid-sized package.

by Qwen | $0.18/M input | $0.7/M output

Qwen3 Coder is the most agentic code model in the Qwen series, available in 30B and 480B MoE variants. It achieves SOTA on SWE-Bench with 256K native context, extendable to 1M tokens.

by Qwen | $1.5/M input | $7.5/M output

Qwen3 Coder 30B A3B Instruct is an efficient MoE coding model with 30B total and 3.3B active parameters, offering strong agentic coding capabilities with 256K context support.

by Qwen | $0.45/M input | $2.25/M output

Qwen3-VL 30B-A3B is a compact mixture-of-experts vision-language model from Alibaba's Qwen team, with 30B total parameters and only 3B active per token for efficient inference. It supports image and text inputs with a 131K context window and delivers strong multimodal performance on benchmarks including MMMU and visual-math evaluations. Capabilities include document and chart understanding, OCR, visual coding (generating HTML/CSS/JS from images), 2D spatial grounding, and GUI agent tasks across desktop and mobile interfaces. The MoE architecture gives it the knowledge breadth of a much larger model while matching the latency and cost profile of a 3B dense model — making it a practical choice for developers who need reliable vision-language capabilities without the compute cost of the 235B flagship variant. Supports tool calling.

by Qwen | $0.2/M input | $0.8/M output

MiniMax Video-01 Director is an AI video generation model that specializes in creating HD videos with precise cinematic camera control. It supports 720p resolution at 25fps and generates clips up to 5 seconds, allowing users to specify camera movements like pans, zooms, and tracking shots through natural language or bracketed commands. The model significantly reduces movement randomness compared to standard video models, enabling more accurate and intentional storytelling.

by MiniMax

QVQ Max is Alibaba's flagship visual reasoning model, built by the Qwen team to combine deep multimodal understanding with rigorous logical inference. Unlike standard vision-language models, QVQ Max is designed to think through what it sees — analyzing charts, diagrams, math problems, and everyday images step by step before responding. It scores 70.3% on MMMU and 71.4% on MathVista (mini), placing it among the top multimodal reasoning models available via API. The model handles text and image inputs across a 131K token context window and supports tool calling for agentic workflows. Ideal for developers building tutoring tools, visual data analysis pipelines, document understanding systems, or any application that requires both image comprehension and structured reasoning.

by Qwen | $1.2/M input | $4.8/M output

Sora 2 is OpenAI's video and audio generation model designed for speed and flexibility, ideal for rapid iteration, concepting, and social media content where quick turnaround matters more than ultra-high fidelity. It generates videos from text prompts or images with synchronized dialogue and sound effects.

by OpenAI

Sora 2 Pro is OpenAI's state-of-the-art, most advanced media generation model that produces higher quality, more polished and stable video results with synced audio. It takes longer to render and costs more, but is best for high-resolution cinematic footage, marketing assets, and production-quality output where visual precision is critical.

by OpenAI

Gemini 2.5 Pro Preview 03-25 is the March 25, 2025 snapshot of Google's Gemini 2.5 Pro, the first model in the 2.5 family. It is a reasoning model that thinks before responding, with a 1M token context window, tool calling, and text, image, audio, and video input. At release it debuted at #1 on the LMArena leaderboard with a score of 1443, about 40 points ahead of the next model. Google reported 86.7% on AIME 2025, 84.0% on GPQA Diamond, and 63.8% on SWE-Bench Verified with a custom agent setup. As a dated preview snapshot its behavior is fixed, which suits workloads that need reproducible output. Later snapshots (05-06, 06-05) superseded it with improved coding performance.

by Google | $1.25/M input | $10/M output

DeepSeek V3-0324 is the March 2025 update to DeepSeek V3, incorporating reinforcement learning techniques from R1 to significantly improve reasoning, coding, and frontend development capabilities. It became the first open-source model to outperform all proprietary non-reasoning models on benchmarks, exceeding GPT-4.5 in math and coding tasks.

by DeepSeek | $0.2/M input | $0.77/M output

Gemini 2.5 Flash is Google's hybrid reasoning model balancing speed, cost, and intelligence with controllable thinking capabilities. It supports up to 1M tokens and excels at summarization, chat applications, and data extraction at scale.

by Google | $0.3/M input | $2.5/M output

Gemini 2.5 Pro is Google's most capable reasoning model with state-of-the-art performance on coding and complex tasks. It features a 1M token context window, advanced multimodal understanding, and Deep Think mode for enhanced reasoning.

by Google | $1.25/M input | $10/M output

OpenAI o1 Pro is a version of o1 with more compute for better responses, designed to think longer and provide the most reliable answers. It's the most expensive model at $150/1M input tokens.

by OpenAI | $150/M input | $600/M output

Mistral Small 3.1 is a 24B multimodal model with 128K context, supporting text and image inputs. It outperforms GPT-4o Mini and Gemma 3 while delivering 150 tokens/second, released under Apache 2.0 for commercial use.

by Mistral AI | $0.35/M input | $0.55/M output

Ideogram 3.0 is a text-to-image generation model from Ideogram AI, built by a team of ex-Google engineers and launched in March 2025. It specializes in photorealistic image generation with industry-leading text rendering — producing accurate, stylized typography within images that competing models like Midjourney and DALL-E 3 struggle to match. The model excels at graphic design tasks including posters, logos, marketing visuals, and layouts with complex or lengthy text compositions. It also supports Style References, allowing up to three reference images to guide output aesthetics for consistent branding across batches. In human evaluations, Ideogram 3.0 achieved the highest ELO ratings against other text-to-image models across diverse prompts covering varied subjects, styles, and composition difficulty. It's a strong fit for developers building design, advertising, or content-generation pipelines where typographic accuracy and prompt adherence are critical.

by Ideogram | $0.0602/image

Gemma 3 12B Instruct is Google's mid-sized open multimodal model supporting text and image input with a 128K token context window. It supports 140+ languages and offers strong performance for single-GPU deployment.

by Google | $0.04/M input | $0.13/M output

Gemma 3 4B Instruct is Google's compact multimodal open model supporting text and images with a 128K token context window. It's optimized for deployment on laptops and edge devices while maintaining strong capabilities.

by Google | $0.05/M input | $0.1/M output

Command A is Cohere's flagship enterprise language model with 111 billion parameters and a 256K token context window, released in March 2025. Built for complex agentic workflows, it leads on tool-use benchmarks including BFCL-v3 and Tau-bench, and performs on par with GPT-4o on MMLU and SQL tasks. It is particularly strong at multi-step tool calling — including knowing when not to invoke a tool, a critical quality for production agents. Supporting 23 languages with 150% higher throughput than Command R+, it's a strong choice for developers building RAG pipelines, autonomous agents, or multilingual enterprise applications.

by Cohere | $2.5/M input | $10/M output

Command A (03-2025) is Cohere's flagship enterprise chat model, built for tool use, retrieval-augmented generation, and agentic workflows across a 256K token context window. Cohere reports it performs on par with or better than GPT-4o and DeepSeek-V3 on agentic enterprise tasks. In Cohere's human evaluations against GPT-4o (November version), it reached win rates of 50.4% on general prompts, 51.4% on business tasks, and 49.6% on code. Cohere also measured serving speeds of up to 156 tokens per second, about 1.75x what it measured for GPT-4o. It supports 23 languages and produces citation-backed answers in RAG settings, which makes it a fit for enterprise assistants, multilingual applications, and multi-step agents that call external tools.

by Cohere | $2.5/M input | $10/M output

Reka Flash 3 is a 21-billion-parameter reasoning model developed by Reka AI, designed as a compact but capable general-purpose LLM. It excels at chat, coding, instruction following, and function calling. The model uses chain-of-thought reasoning via explicit thinking tags, and supports a "budget forcing" mechanism that lets you cap reasoning steps to control latency. It offers a 130K-token context window and is text-only (no image input). Reka AI positions it as competitive with OpenAI's o1-mini while being significantly smaller. It scores 65.0 on MMLU-Pro — modest for knowledge-heavy tasks, so pairing it with search or retrieval is recommended. It's primarily English-focused. Priced at $0.20 per million input tokens and $0.80 per million output tokens via the Reka API, it's a cost-effective option for developers who need solid reasoning at low cost.

by Reka AI | $0.1/M input | $0.2/M output

Gemma 3 27B Instruct is Google's most capable single-GPU open model with multimodal support, 128K context, and 140+ language support. It outperforms many larger models and offers state-of-the-art open-weight performance.

by Google | $0.08/M input | $0.16/M output

Skyfall 36B v2 is a 36-billion parameter model that upscales Mistral Small 2501 with specialized training for creativity, roleplay, and coherent storytelling. Users report it rivals or exceeds 70B parameter models in creative writing quality while remaining accessible for local deployment with strong chain-of-thought reasoning and tool use capabilities. It features a 32K token context window and supports Mistral v7 Tekken, Metharme, and Alpaca chat templates.

Sonar Deep Research is Perplexity's expert-level research model designed for exhaustive multi-step retrieval, synthesizing hundreds of sources into comprehensive reports. It autonomously searches, reads, and evaluates sources while refining its approach for in-depth analysis across domains like finance, technology, and health. Ideal for detailed market analyses, literature reviews, and projects requiring synthesis of multiple information sources.

by Perplexity | $2/M input | $8/M output

Sonar Pro is Perplexity's advanced search model with a 200K token context window, delivering 2x more citations and search results than standard Sonar for complex queries. It handles in-depth, multi-step queries with enhanced content understanding and supports longer, more nuanced follow-up conversations. Best for enterprise applications requiring deeper research and comprehensive source attribution.

by Perplexity | $3/M input | $15/M output

Sonar Reasoning Pro is a premier reasoning model powered by DeepSeek R1 with Chain of Thought (CoT), designed for complex multi-step analysis and logical problem-solving. It excels at tasks requiring step-by-step thinking, strict instruction adherence, and information synthesis across sources with built-in web search. Ranked among the top models in Search Arena evaluations, statistically tied with Gemini-2.5-Pro-Grounding.

by Perplexity | $2/M input | $8/M output

QwQ Plus is a proprietary reasoning model from Alibaba's Qwen team, serving as the hosted API counterpart to the open-weight QwQ-32B release. Like QwQ-32B, it uses reinforcement learning to develop extended chain-of-thought reasoning, excelling at math competition problems, scientific reasoning, and complex coding tasks. QwQ-32B achieved 79.5% on AIME 2024, 90.6% on MATH-500, and 63.4% on LiveCodeBench — rivaling much larger models. QwQ Plus exposes these capabilities through a managed API endpoint with a 131K token context window and tool call support. Best suited for developers building applications that require step-by-step mathematical reasoning, algorithmic problem-solving, or multi-step logical inference.

by Qwen | $0.8/M input | $2.4/M output

Kling 2.0 Master is Kuaishou's flagship model from the 2.0 generation, delivering 1080p cinema-grade video with 3D spatiotemporal joint attention for realistic motion and physics simulation. It marked a major leap in visual realism and semantic understanding, supporting up to 5-second videos at 24fps with a multi-elements editor for flexible scene control.

by Kling

Grok 3 is xAI's flagship model launched February 2025, trained with 10x more compute on the Colossus supercluster with 200,000 GPUs. It features advanced reasoning through reinforcement learning, deep domain knowledge in finance/healthcare/law/science, and a 131K token context window.

by xAI | $3/M input | $15/M output

Grok 3 Fast is a latency-optimized variant of Grok 3 using the same underlying model but served on faster infrastructure. It delivers quicker response times for latency-sensitive applications while maintaining equivalent reasoning quality and 131K context window.

by xAI | $5/M input | $25/M output

Grok 3 Mini is a lightweight, cost-efficient reasoning model that thinks before responding, ideal for logic-based tasks that don't require deep domain knowledge. It features configurable reasoning effort and exposes accessible thinking traces for transparency.

by xAI | $0.3/M input | $0.5/M output

Grok 3 Mini Fast is the speed-optimized variant of Grok 3 Mini, running on faster infrastructure for significantly quicker response times. It provides identical reasoning quality to Grok 3 Mini but is designed for latency-sensitive applications.

by xAI | $0.6/M input | $4/M output

Mistral Saba is a 24B parameter regional model trained for Arabic and South Asian languages including Tamil and Malayalam. It outperforms models 5x its size on Arabic benchmarks while providing culturally relevant responses.

by Mistral AI | $0.2/M input | $0.6/M output

OpenAI o3 Mini High is a higher-intelligence version of o3-mini that takes longer to generate more accurate responses. It uses high reasoning effort for complex STEM and coding tasks.

by OpenAI | $1.1/M input | $4.4/M output

Aion RP 1.0 8B is an uncensored roleplay and creative writing model from AionLabs, fine-tuned from the Llama 3.1 8B base model rather than an instruct variant. This base-model approach is designed to produce more natural and varied writing. It ranks highest in the character evaluation portion of RPBench-Auto, a roleplaying-specific benchmark derived from Arena-Hard-Auto where LLMs evaluate each other's responses. The model supports the full 131K context window and multi-turn conversations. Best suited for character-driven chat applications, interactive storytelling, and persona-consistent dialogue. A recommended temperature of 0.7 is advised, as higher values can degrade output quality.

by Aion Labs | $0.8/M input | $1.6/M output

Qwen2.5 7B Instruct 1M is a 7-billion-parameter instruction-tuned chat model from Alibaba's Qwen team, released in January 2025 as the long-context variant of Qwen2.5 7B Instruct in the Qwen2.5-1M series. It handles contexts up to 1 million tokens, enough to fit entire codebases, books, or large document collections in a single request. In Qwen's 1M-token passkey retrieval test it finds hidden information with near-perfect accuracy (minor errors reported for the 7B size), and on RULER, LV-Eval, and LongBench-Chat it outperforms its 128K predecessor, especially on sequences beyond 64K tokens. On short-context academic benchmarks it performs about the same as the 128K version, which Qwen reports as comparable to GPT-4o-mini, while supporting a context eight times longer. A low-cost option for long-document question answering, retrieval, and analysis over very large inputs.

by Qwen | $0.37/M input | $1.47/M output

Sonar is Perplexity's lightweight, cost-effective search model built on Llama 3.3 70B, optimized for speed (1200 tokens/second) and quick factual queries. It provides real-time web search with grounding and citations, ideal for simple Q&A and straightforward integrations. Best for everyday use cases where fast, accurate answers are needed without complex reasoning.

by Perplexity | $1/M input | $1/M output

Qwen-Omni Turbo is Alibaba's cost-optimized omnimodal API model, built to process text, image, audio, and video inputs and return text responses in a single unified interface. It is the lighter, faster tier in the Qwen-Omni family, designed for developers who need full multimodal coverage at lower latency and cost than the flagship Qwen-Omni model. Audio files up to 40 seconds and video files up to 150 MB are supported, spanning common formats such as MP3, WAV, MP4, and MOV. The model handles tool calling natively and is accessible via an OpenAI-compatible API. Best suited for developers building applications that need to reason across mixed media inputs — such as audio transcription pipelines, video understanding workflows, or multimodal chatbots — where throughput and cost efficiency matter.

by Qwen | $0.07/M input | $0.27/M output

Moonshot V1 8K Vision Preview is a multimodal variant of Moonshot AI's V1 model that accepts both image and text inputs within an 8,000-token context window. It can interpret screenshots, charts, UI mockups, and photos, returning text-based analysis. This makes it useful for tasks like image captioning, visual Q&A, and lightweight document understanding where the source material includes visual elements. As a preview model, it may see changes before a stable release. The API follows the OpenAI-compatible content array format with image_url blocks, making integration straightforward for developers already using similar patterns.

by Moonshot AI | $0.2/M input | $2/M output

Moonshot V1 32K Vision Preview is a multimodal model from Moonshot AI that processes both images and text within a 32,000-token context window. It extends the base 32K model with the ability to interpret visual inputs — including screenshots, diagrams, charts, and scanned documents — and return text-based responses. This is useful for workflows that combine visual context with moderate-length text, such as analyzing annotated documents or explaining UI designs. As a preview release, the vision capabilities may evolve. The API accepts the standard OpenAI-compatible content array format for multimodal inputs.

by Moonshot AI | $1/M input | $3/M output

Moonshot V1 128K Vision Preview is Moonshot AI's largest-context multimodal model in the V1 series, supporting both image and text inputs within a 128,000-token context window. It combines the long-context strength of the 128K text model with visual understanding capabilities. This makes it well-suited for processing large multimodal documents — think lengthy reports with embedded charts, multi-page scanned PDFs, or extensive UI review sessions. As a preview model, vision features may be refined over time. The API uses the standard OpenAI-compatible format for multimodal content, making it a drop-in addition to existing workflows.

by Moonshot AI | $2/M input | $5/M output

MiniMax-01 is a 456B parameter foundation model (45.9B activated) using a hybrid Lightning Attention + MoE architecture, achieving top-tier performance on reasoning, math, and coding benchmarks. It supports up to 4 million tokens of context, making it especially strong for long-context tasks and AI agent applications.

by MiniMax | $0.2/M input | $1.1/M output

Phi-4 is a 14B parameter small language model from Microsoft that excels at complex reasoning tasks, especially mathematics, outperforming many larger models on math competition benchmarks while being efficient enough for edge deployment.

by Microsoft | $0.07/M input | $0.14/M output

Vidu 2.0 is an AI video generation model by ShengShu Technology that creates high-quality videos from text or images, supporting resolutions up to 1080p. It offers smoother motion, better frame consistency, and start/end frame control compared to its predecessor. The model is significantly faster and more affordable.

by Vidu

Qwen-MT Plus is a specialized machine translation model from Alibaba's Qwen team, purpose-built for high-quality text translation across 92 languages covering over 95% of the world's population. Unlike general-purpose language models, Qwen-MT Plus is fine-tuned specifically for translation tasks, offering term intervention, domain prompting, and translation memory features that give developers fine-grained control over output. It supports translation between major languages including Chinese, English, Japanese, Korean, French, Spanish, German, Arabic, Thai, Indonesian, and Vietnamese. Best suited for developers building multilingual applications, content localization pipelines, or customer-facing translation features where accuracy, terminology consistency, and domain fidelity matter more than general conversational ability.

by Qwen | $2.46/M input | $7.37/M output

Qwen-MT Turbo is a fast, cost-effective machine translation model from Alibaba's Qwen team, designed for high-volume text translation across 92 languages. As the Turbo tier of the Qwen-MT family, it trades some of the output fidelity of Qwen-MT Plus for significantly lower cost and faster throughput — making it the practical choice for latency-sensitive or budget-constrained translation workflows. Like its sibling, it supports term intervention, domain prompting, and translation memory, giving developers control over terminology and style. Best suited for developers building high-volume localization pipelines, real-time translation features, or cost-sensitive multilingual applications where speed and price efficiency matter more than maximum output quality.

by Qwen | $0.16/M input | $0.49/M output

Qwen Plus Character is the role-playing model in Alibaba's Qwen series, offered through Alibaba Cloud Model Studio with an OpenAI-compatible chat API. It is tuned for staying in character across a conversation. Alibaba's documentation highlights following predefined character instructions, advancing the dialogue, and showing active listening and empathy, with support for detailed persona definitions set through the system prompt. Compared with the general-purpose Qwen Plus, it trades broad task coverage for character consistency and natural conversational flow. It handles text input and output only and does not support function calling, so it is not a fit for tool-using agents. Alibaba lists virtual social apps, game NPCs, IP character replication, and smart hardware assistants as target use cases. A Japanese-tuned variant (qwen-plus-character-ja) is also available.

by Qwen | $0.5/M input | $1.4/M output

OpenAI o3 Mini is a cost-efficient reasoning model specialized for STEM domains requiring precision and speed. It features three reasoning effort levels (low, medium, high) and supports function calling.

by OpenAI | $1.21/M input | $4.84/M output

Llama 3.3 Euryale 70B v2.3 is the latest in Sao10K's Euryale series, built on Llama 3.3 Instruct with a 131K context window and 16K output limit. It's a direct successor to v2.2, trained without LoRA extraction for more robust creative roleplay and storywriting performance.

Llama 3.3 Euryale 70B v2.3 is a creative writing and roleplay finetune of Meta's Llama 3.3 70B Instruct by Sao10K, an independent model creator whose Euryale line is their flagship series for these tasks. v2.3 is the successor to Euryale v2.2. Unlike earlier versions, it is trained entirely on top of Llama 3.3 Instruct rather than LoRA-extracted, and the creator describes it as a direct replacement with improved performance. Its training data mixes roleplay and creative content with instruction data, so it handles storytelling, interactive dialogue, and general instruction following. The model uses the Llama 3 Instruct prompt format, and the creator recommends a temperature around 1.1 with min_p 0.1. Through the API it offers a 131K context window with up to 16.4K output tokens, a fit for long multi-turn roleplay and story sessions.

Kling 1.6 Standard is Kuaishou's accessible 720p AI video model released in December 2024, offering a 195% improvement over Kling 1.5 in image-to-video quality. It provides fast, consistent video generation with enhanced prompt adherence and natural motion, ideal for beginners and creators needing quick social media content.

by Kling

Google Veo 2 is Google DeepMind's video generation model that creates 5-second, 720p-4K resolution videos from text or image prompts with realistic physics simulation and cinematic quality. It excels at following complex instructions, simulating real-world physics, and supporting diverse visual styles without native audio generation.

by Google

Command R7B is Cohere's smallest and fastest model in the R series, with 7 billion parameters and a 128K token context window. Despite its compact size, it ranked first among similarly-sized open-weights models on the HuggingFace Open LLM Leaderboard, leading across IFEval, BBH, GPQA, MuSR, and MMLU. It supports native tool use, multi-step agentic workflows, and RAG across 23 languages, with particular strength in code tasks including SQL and code translation. For API developers, it's the best option when latency and cost are priorities and a full-scale model isn't required.

by Cohere | $0.04/M input | $0.15/M output

72B Qwen2.5 Kunou v1 is a roleplay and creative writing model by Sao10K, the independent creator behind the Euryale and Stheno series, finetuned from Qwen2.5 72B Instruct. Sao10K describes it as a generalist focused on roleplay and creative instruction tasks, and as something of a successor to Llama 3 Euryale v2.2, trained on a cleaned-up version of the datasets used for Euryale and Stheno. The model uses the ChatML prompt format. The creator notes that system prompts have a strong effect on its behavior, and recommends a temperature of 1.1 with min_p 0.1. It is the largest model in the Kunou series (14B and 32B versions also exist) and suits developers who want Euryale-style roleplay and storytelling on a Qwen base instead of Llama.

OpenAI o1 is a reasoning model that thinks before answering using chain-of-thought, excelling at complex science and mathematics tasks. It was the first in OpenAI's "o" series designed for step-by-step logical reasoning.

by OpenAI | $15/M input | $60/M output

Amazon Nova Lite is a very low-cost, lightning-fast multimodal model that processes text, images, and video inputs to generate text output. It supports up to 300K input tokens and can analyze multiple images or up to 30 minutes of video in a single request. Ideal for real-time customer interactions, document analysis, and visual question-answering tasks.

by Amazon | $0.06/M input | $0.24/M output

Amazon Nova Micro is a text-only model that delivers the lowest latency responses at the lowest cost in the Nova family. With a 128K token context window, it excels at text summarization, translation, content classification, interactive chat, and basic coding tasks. It's the fastest and most economical option when multimodal capabilities aren't needed.

by Amazon | $0.04/M input | $0.14/M output

Amazon Nova Pro is a highly capable multimodal model offering the best combination of accuracy, speed, and cost for a wide range of tasks. It supports up to 300K input tokens, excels at video summarization, financial document analysis, agentic workflows, and can process code bases with over 15,000 lines of code. It also serves as a teacher model for distilling custom variants of Nova Micro and Lite.

by Amazon | $0.8/M input | $3.2/M output

GPT-4o 2024-11-20 is a November 2024 snapshot of GPT-4o providing the latest improvements at that time. It's useful for applications requiring locked model behavior.

by OpenAI | $2.5/M input | $10/M output

Mistral Large 2 (24.07) is a 123B parameter model with 128K context, significantly upgraded for long context understanding and function calling. It delivers top-tier performance for enterprise use cases including knowledge exploration and automation.

by Mistral AI | $2/M input | $6/M output

Qwen 2.5 Coder 32B Instruct is a code-specialized model matching GPT-4o's coding capabilities, supporting 40+ programming languages. It excels in code generation, repair, and reasoning with 128K context support.

by Qwen | $0.66/M input | $1/M output

UnslopNemo 12B is a 12-billion parameter model where TheDrummer removed repetitive patterns ('slop') from roughly 90% of the roleplay training dataset to make outputs more expressive and varied. It's designed for adventure writing and roleplay scenarios with a 1M token context window. The model aims to generate more natural, less formulaic creative content compared to standard fine-tuned models.

Qwen Turbo is a fast, cost-effective API model with up to 1M context length, ideal for simple tasks requiring quick responses. It supports multiple languages and offers flexible tiered pricing.

by Qwen | $0.05/M input | $0.2/M output

Magnum v4 72B is a 72-billion parameter creative writing and conversational model developed by Anthracite, fine-tuned on top of Qwen2.5-72B-Instruct. It was specifically trained to replicate the prose quality of Claude 3 Sonnet and Opus, making it one of the most popular open-weight models for narrative generation. The model excels at creative writing, interactive storytelling, roleplay, and character-driven dialogue. It actively drives narratives forward while maintaining consistent character personas across extended conversations. Multi-language support covers English, French, German, Spanish, Chinese, Japanese, and more. Magnum v4 72B supports up to 32,768 tokens of context and uses the ChatML prompt format. It's a strong choice for developers building applications where engaging, human-like prose matters more than raw benchmark performance.