Rate limits

30 requests per minute

The default account guardrail is 30 requests per UTC minute across keys. A rejected request returns 429 and Retry-After: 60. The limit protects shared capacity and may be adjusted through documented account policy.

• Use bounded concurrency.
• Back off on 429 and 503.
• Do not create extra accounts to evade limits.

Payload bounds

Requests have a bounded body and max output. max_tokens and max_completion_tokens must be integers from 1 through 32,768; a request cannot reserve unbounded credit.