Wall Street finally has a price for embodied AI

For four years the humanoid robot industry has run on promises. On Wednesday it got a number: $51B.

That is what public markets paid for Unitree on its first day of trading, a 460% close that made its founder roughly $13B richer before lunch and gave an entire category its first real comp. He spent the afternoon on a Beijing stage explaining that humanoids are still two to ten years from their ChatGPT moment. Investors bid anyway.

Here is the part nobody put in a press release. In the same 72 hours, three labs shipped the other half of the machine independently. DeepSeek made vision cheap enough to run in an agent loop. Meta previewed open weights on a model with a robot brain attached. OpenAI walked an agent into your private message archive. Not one of those is a chatbot feature.

In today’s AI news:

• Unitree’s debut sets the first public price for embodied AI

• DeepSeek makes machine vision cheap enough to run in a loop

• Meta previews open weights on a model that drives robots

• OpenAI walks an agent into your iMessage archive

News: Unitree closed 460% above its 150.80 yuan offer price in its Shanghai debut, the first humanoid robot maker to trade on a mainland Chinese exchange. The close valued the company near $51B and lifted founder Wang Xingxing’s stake to roughly $16B. Hours later he told a Beijing conference that humanoids are two to ten years from their ChatGPT moment.

Details:

• Shares opened at 1,100 yuan, up as much as 629% intraday, before closing at 845. The IPO raised $904M at a pre-trading valuation near $9B.

• Only 10% of enlarged share capital hit the float. Allocations to DeepSeek, Tencent, Alibaba, Ant Group and Meituan carry 12 to 36 month lockups.

• The fundamentals are real: 2025 revenue of 1.7B yuan against 393M a year earlier, 278M yuan net profit, gross margin above 60%.

• Unitree shipped 5,500+ humanoids last year, more than any manufacturer globally, with 33,000+ quadrupeds sold cumulatively.

• Wang defines the breakthrough as a robot completing 80% of voice commands in a home it has never seen.

Why It Matters: Until Wednesday nobody knew what public markets would pay for embodied AI, and now 30 to 50 Chinese robotics companies are queued for listings that will price off this comp. The prospectus carries the warning label: Unitree’s buyers are still overwhelmingly universities and research institutes, so the market just assigned $51B to a company selling into R&D budgets. Treat it as a financing window rather than proof of demand. The capital is available now and the customers arrive later, which means raise on this comp and build for the one that comes after it.

News: DeepSeek has put DeepSeek-V4-Flash-Vision-Exp live on its API platform, the lab’s first hosted multimodal model. It matches V4-Flash on text work including agents, reasoning and world knowledge, and on agent benchmarks requiring visual understanding it lands close to Opus-4.8. A new Files API shipped with it, closing a gap developers have been filing issues about since January 2025.

Details:

• Set model=’deepseek-v4-flash-vision-exp’ to access it. DeepSeek Harness 0.1.1 landed the same day with out-of-the-box support.

• Images cap at 384 tokens each at standard V4-Flash pricing. A 2000x2000 image and a 5000x5000 image cost identically after resizing.

• Three input paths: inline base64, a public URL, or a file_id. Requests accept up to 600 images against a 48 MiB body limit.

• The Files API is free and gives every key 25 GiB and 10,000 stored files, with expiry from one hour to permanent.

• It runs across Chat Completions, the Responses API, and the Anthropic-compatible /messages endpoint.

Why It Matters: Every benchmark DeepSeek chose to cite is an agent benchmark, which tells you exactly who this was built for. That 384-token ceiling turns screenshot-driven loops from a line item into a rounding error, and it makes browser agents, GUI automation and document pipelines viable at a scale most teams have been priced out of entirely. The Anthropic-compatible endpoint is the sneaky part: DeepSeek is not asking you to migrate, it is asking you to point an existing harness somewhere cheaper.

News: Meta Superintelligence Labs published the multimodal side of Muse Spark 1.2 on August 20, explicitly framed as arriving ahead of the open-weights release. Buried under the benchmark charts is a robotics variant that plans and orchestrates physical tasks. Meta also published several evaluations it loses, including one it loses to its own predecessor.

Details:

• A specialized variant runs as robot brain in a two-level stack: a planner handles scene understanding and decomposes goals, while a Muse Spark VLA policy executes each subtask.

• The tool-use delta is the number worth copying down: multimodal intelligence climbs from 59.8 without tools to 72.0 with them. Version 1.1 starts higher at 60.2 and only reaches 69.1.

• On ZeroBench it hits 54.0%, just behind GPT-5.6 Sol at 54.6% and well ahead of Opus 5 at 47.5%. On CharXiv Reasoning it scores 87.6%, behind Muse Spark 1.1 at 88.4%.

• On Design Arena as of August 18, it takes first in Video to Website at 1279 Elo, second on Image to Website, third on Image to Frontend.

• Meta notes the low-level robot actions come from an experimental smaller Muse model and that grasps in the simulated demos are simplified.

Why It Matters: A model that is marginally worse cold and meaningfully better with tools is not a weaker model, it is one built for agent loops instead of leaderboards, and Meta published the losses to make that case. WildArtifactBench is the tell: 10 tasks from an internal eval scoring agents by win rate against a baseline, no ground truth required, plotted against cost per task. Meta is not conceding the benchmark, it is proposing a new one. If open weights land with the VLA variant intact, a robotics stack that cost a seed round in January becomes a weekend project. Get your eval harness written now so you can move the day it drops.

News: OpenAI shipped an Apple Messages plugin for the ChatGPT Mac app on August 20, letting the assistant read, search, draft and send across iMessage, SMS and RCS conversations stored on your machine. It runs inside ChatGPT Work and Codex, not regular ChatGPT chats, and it is available on every plan including free. This is the deepest OpenAI has reached into a native Apple app, and it landed six weeks after Apple sued the company.

Details:

• The plugin requires Full Disk Access, Contacts and Automation permissions and ships only in the Apple silicon build. Intel Macs are excluded.

• Sending is gated by default. ChatGPT surfaces the message and recipients for approval, with “Allow once” or “Always allow sending to this chat”.

• OpenAI warns against persistent approval because it removes the last human checkpoint, and flags a known issue with tasks that disable approval prompts.

• Processing happens locally and OpenAI says it builds no general index of your messages.

• It follows Computer History expanding to Pro, Business and Enterprise users in the EEA, Switzerland and the UK.

Why It Matters: Placement is the whole story. OpenAI put Messages inside Codex and ChatGPT Work rather than regular chat, positioning your message history as agent context instead of chatbot input, and those are different products with different risk surfaces. The consent math is worth sitting with: grant access and every person who has ever texted you is opted in without being asked, which no dashboard on your side resolves. Apple sued OpenAI in July over alleged trade secret theft, and OpenAI is now shipping deeper into Apple’s apps while asking a judge to dismiss. If you are building anything that touches local user data, read OpenAI’s approval-gate design before you write your own. They shipped the pattern regulators will benchmark against.

• 📁 DeepSeek Files API is free, gives you 25 GiB and 10,000 stored files per key, and lets you upload an image once and reference it by file_id forever.

• 🧪 WildArtifactBench opens 10 tasks from Meta’s internal agent eval, scored by win rate against a baseline with no ground-truth deliverable required.

• ⚙️ DeepSeek Harness 0.1.1 shipped same-day support for the new vision model, so agent frameworks work with images without custom plumbing.

• 🎨 Design Arena ranks models head-to-head on Image to Website, Video to Website and Image to Frontend, which is the closest public proxy for visual coding quality.

• 🔌 Meta Model API is where Muse Spark 1.2 lives today at $1.25 in and $4.25 out per million tokens, with a 1M-token context window.

• Washington added foreign-made humanoid and quadruped robots to the FCC covered list, restricting imports. The US accounted for 13.3% of Unitree’s 2025 revenue, and the company says current models hold approvals while future ones may not.

• Tesla is putting the steering-wheel-free Cybercab on Austin roads as soon as this month, starting with employee rides. Tesla is self-certifying rather than taking the NHTSA exemption route Zoox used.

• Amazon is expanding Prime Air sixfold to nearly 500 US cities and towns by year-end, up from 11 sites across 10 metros, with 30-minute delivery on packages under five pounds.

• Three ex-SpaceX engineers opened an autonomous steel fabrication plant in Cincinnati on a $15M seed, welding the skids under AI data centers. Robotic arms hit 70% arc-on time against 10 to 12% for humans.

• SoftBank invested $200M in Swiss startup Gravis Robotics, which retrofits construction machinery into autonomous equipment.

• Chery’s robotics arm AiMOGA is preparing an IPO, targeting 10K robot deliveries next year.