Rate and concurrency limits

Read actual limits; execution budgets and request frequency differ.

Anonymous catalog reads, cheap queries, user/Key execution RPM, concurrency and costly-task budgets have separate limits. Scope and model counts do not imply quotas; historical unlimited PAYG values need a documented migration when tightened.

Respect applicable Retry-After and bound polling with backoff. Disconnecting releases concurrency slots without erasing late measurement. Dependency outages stop new costly work by capability while preserving queries, revocation and paid fulfillment recovery. A rate-limit response does not prove that earlier work did not execute.

Pool-3 text has a separate prepaid resource contract: output at most 1,048,576 tokens, subject to a lower model cap; n=1; no max_completion_tokens; finite, nonnegative integer thinking.budget_tokens no greater than effective max_tokens. The full reservation estimate must be an exact integer at most 2^53−1. These are execution bounds, not per-minute quotas or a promise that every provider supports them. Member entitlement limits keep their own semantics.

Go's candidate text-body ceiling is 32 MiB. Apply the lower of Go, nginx/proxy and upstream limits, including any deployment-specific restriction. A 413 can occur before admission; native /v1 errors and platform errors use their respective envelopes. Production enablement and actual limits need a deployment receipt.