Model leaderboard

Methodology

Only our own traffic

Every request through the FreeTheAI gateway is counted once into time buckets: 5-minute, hourly, and daily. The 7-day and 30-day views use whole UTC days. No vendor numbers, benchmarks, or votes are used. Growth compares each model with the window just before it, using only finished hours for the 24-hour view and finished days for the others, so a day still in progress never looks like a drop. A model with no traffic in the earlier window shows no growth figure.

How rows are ranked

Models are ranked by tokens served (input, cache reads, and output); when two are tied, the one with more requests ranks higher. Share is a model's part of all tokens served in the window. Only models listed in the catalog appear. The headline cards need at least 20 requests in the window, so a handful of lucky requests cannot take a record.

What the numbers mean

Success rate counts requests the provider answered or failed; a caller's own bad request never counts against a model. Latency is the full request time. Speed is measured on streamed replies only: output tokens per second from the first generated token to the last, using the provider's own token count, and the board shows the median reply. Replies that were not streamed, or whose count had to be estimated, are left out.