OpenRouter and a16z "State of AI" report finds reasoning models above half of all tokens
The "State of AI" report (published December 2025) analyses more than 100 trillion tokens of real LLM interactions on OpenRouter over a rolling 13-month window ending November 2025 and finds that the…
- Date
- December 2025
- Who
- OpenRouter, Andreessen Horowitz (a16z)
- People
- Malika Aubakirova, Anjney Midha, Alex Atallah, Chris Clark, Justin Summerville
- Confidence
- Medium (one usage platform with a developer-skewed user base)
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): OpenRouter, Andreessen Horowitz (a16z) · People: Malika Aubakirova, Anjney Midha, Alex Atallah, Chris Clark, Justin Summerville · Confidence: Medium (one usage platform with a developer-skewed user base) The "State of AI" report (published December 2025) analyses more than 100 trillion tokens of real LLM interactions on OpenRouter over a rolling 13-month window ending November 2025 and finds that the share of tokens routed through reasoning-optimised models climbed from negligible in early 2025 to above 50%; programming grew from about 11% to over 50% of token volume, and tool calling and multi-step "agentic inference" grew steadily (OpenRouter). It is the only large usage measurement of reasoning adoption found for this chapter and closes an earlier Backlog item. My inference is that the users are developers who pick models, so the figures describe API usage on one platform and do not cover ChatGPT or the general population, and that because hidden thinking tokens are billed as output (B08-06) token share overstates request share. The report says category-level data begins only in May 2025. See also B18. Sources: OpenRouter, State of AI