Rate limits
The rate-limit layers and their headers.
Limits stack in independent layers; hitting any one of them returns 429 with the same shape:
Every response (not just 429s) carries your current standing:
Always honor Retry-After (seconds) before retrying — it's the one header guaranteed to stay on v1.
Limits are per account, by call class (read/write/heavy — generation-style calls), not per key:
issuing more keys doesn't raise your ceiling. Plan-level daily/monthly quotas (where a product has them)
return 503 service_unavailable instead of 429 if they can't be checked — rare, and only for API-key/OAuth
traffic, never for the product's own web app.
Next: Errors · Idempotency