100 Trillion Tokens Later

2025 was a big year for AI. We got reasoning models, Chinese open-source labs landed on the scene, and programming surged as a dominant use case. A lot has changed.

I just finished digging through OpenRouter’s State of AI report, which analyzes usage across 100 trillion tokens. I pulled out the most interesting data below, but in general the core trends are:

• Total usage is waaaaay up

• The gap between open-source and closed-source models is narrowing

• Programming was not the number one use case

• Specific models have better retention than I would’ve guessed

• The split stayed consistent

• But it went from 0% Chinese to mostly Chinese (Meta falls off, Mistral is meh)

• The fact that it was above 30% for portions of the year is wild

• Weekly token usage grew roughly 70× year-over-year, from ~100 billion to over 7 trillion tokens per week, with closed models driving most volume

• Open-source models quietly capturing ~¼ of total usage by the end of the period.

• One of the most surprising graphs, I would’ve guessed that the top use case would’ve been programming

• Closed-source still dominates programming usage, but open source grew from ~2% to ~15% over the year (5–7× increase, ~500% growth), peaking at ~30% of programming tokens at its high point.

• Reasoning tokens were introduced with OpenAI’s o1, followed by DeepSeek’s R1

• Reasoning has led to much better performance, but at the cost of burning many more tokens (more on this ratio later)

• Being proficient at calling tools is now table stakes for modern LLMs

• A larger green portion = the more we’re using LLMs beyond a search replacement

• OpenAI and Anthropic were the first to ship models that could reliably call tools, with Claude Sonnet 3.5 clearly ahead for an extended period.

• Larger prompt tokens mean users are probably sending more complex prompts, expecting models to be able to handle such complexity

• Everything is flat except coding

• In programming prompts, users are more likely to paste large pieces of documentation, code examples, or files

• Average sequence length has more than tripled over ~12–14 months, increasing from ~1,700–1,900 tokens in early/late 2024 to over 5,400 tokens by December 2025.

• The increase in sequence length is driven mostly by programming LLM calls

• 2025 is the year LLM coding really started to work and work consistently

• Most LLM users churn quickly, testing and moving on

• A small number of early cohorts show durable retention. These cohorts emerge when a model achieves strong workload–model fit. i.e., they had tasks that were finally achievable at a reasonable price–performance balance.

• The researchers call this the Cinderella Glass Slipper effect: when a model is the first to solve a previously unmet, high-value workload, users lock in.

• Once embedded in workflows and infrastructure, there isn’t a need to switch. Switching costs are often much higher than just replacing the model name string. If it’s not broken, don’t fix it.