Rankings

How these rankings are measured

What is measured

Every request routed through this gateway, counted from daily per-model totals. Tokens are prompt plus completion plus reasoning; requests are calls, successful or not. The totals carry no user dimension at all — they are aggregated with that column summed away before anything here can read them, and the tables they come from have no user_id to join on.

Selection period

The trailing 30 days. The change beside each model compares that window against the 30 days immediately before it. A model with no traffic in the earlier window is marked new rather than shown as an infinite rise.

Why a model can be missing a name

A model that has been disabled or removed from the catalog still has traffic in the window, and it is still listed — by its bare id, grouped under “Unattributed”. Dropping it would make the percentages here disagree with the totals above them.

Tokens and requests are different rankings

Audio, image and transcription models report no tokens at all, so ranking by tokens excludes them rather than listing them at zero. Switch to requests to see them. Token volume is also not a measure of preference — models differ in how verbosely they answer and how they tokenize.

What these rankings do not establish

Adoption, not quality. Nothing here measures accuracy, reasoning or benchmark performance — those are on the benchmarks page, measured by third parties. And this is traffic through one router, not the market: a model heavily used on its own provider’s API can be quiet here.