Window. The page requests the current 30-day aggregate. The snapshot timestamp is displayed above. Counts come from daily usage rows belonging to listed developers and exclude rows held from public rankings.

Normalization. Date-suffixed and gateway-prefixed raw model identifiers are merged into normalized model families before token buckets and distinct developers are counted.

Sample floors and coverage. A model must have at least 2 distinct participating developers and 100,000 combined uncached-input, cache-write, and cache-read tokens. The API returns at most 60 models ranked by total tokens, so ratio leaders are leaders among returned rows meeting the floor. These rules reduce one-person and tiny-denominator artifacts; they do not make the cohort statistically representative.

Reporting limits. Token categories are taken from agent/provider usage records and may not be reported consistently by every integration. A zero cache-read value can mean no observed cache reuse or unavailable cache telemetry; this aggregate cannot distinguish those cases.

Cohort limit. Public participation is opt-in. Results describe this cohort's coding sessions, not market share or a survey of all AI developers. They should be treated as directional observed behavior.