Annual report · January to November 2026, updated as the year runs
Which models, labs and providers carried production traffic this year. Measured from real token volume routed through LLM Gateway, not from benchmarks or surveys.
The short version of 2026.
Total tokens processed each month, with the model that led it.
Ranked by tokens across every provider that serves the model.
Token volume grouped by the lab that built the model.
Who actually served the tokens, including third-party hosts.
Growth in token volume over the most recent months of the period, against the months before them.
+3,365%
Qwen3.8 Flash901.7M to 31.2B tokens
+1,258%
GLM-5.3 Flash13.4B to 182.6B tokens
+883%
GLM-5.39.4B to 92.1B tokens
+202%
GPT-5.6 Terra7.4B to 22.4B tokens
+35%
MiniMax M311.1B to 14.9B tokens
How to read the numbers.
The report counts tokens and requests routed through LLM Gateway in 2026, in UTC, for models in the public catalogue. A model served by several providers is counted once in the model ranking and split by host in the provider ranking.
Usage shows where production workloads run. It is not a quality score: a model can top a benchmark and still see little real traffic. Gateway traffic also leans towards developers who route across providers, so read shares as a view of that audience rather than of the whole market.
One OpenAI-compatible endpoint, one key, automatic failover between providers.
Other years: 2025