Rate limits and retries¶
Treat 429 and transient 5xx responses as conditions that require bounded client handling, not as a promise of a fixed retry header, limit, or error schema. Preserve the request ID, endpoint, model or resource ID, timestamp, and redacted error detail.
Retry only requests that are safe to repeat. For writes or asynchronous calls, first query the known task, run, or resource and use a documented idempotency mechanism when one is available.
Last updated on