Daily limits
Every metered key has two independent daily caps, reset at midnight UTC:
Both are checked before the request runs, so a blocked request is never billed. Caps are configured per key — ask us if your workload needs more. Partner (unmetered) keys have no daily caps.
Request size limits: at most 1,000 messages per request, and
max_tokens may not exceed the model’s context length.
Error envelope
Errors follow the OpenAI format, so existing SDK error handling works:limit object appears only on 429 responses — cap is the configured cap (requests, or spend in micro-USD).
Status codes
Handling failures
- Retries:
5xxresponses are safe to retry — failed calls are never charged. Use exponential backoff; don’t retry400/401. 429: back off until midnight UTC or switch to a cheaper model; thelimitobject tells you which cap you hit.- Support: every response includes an
X-Request-Idheader. Include it in any report — we can trace the exact request.