Rate limits

Limits apply at two levels: per API key, and per account. Whichever is stricter wins.

Defaults

Set a lower rate_limit_per_minute on any key at creation time to bound a specific integration below the account default.

Headers

Every response carries the current window state:

X-RateLimit-Limit: 30
X-RateLimit-Remaining: 24
X-RateLimit-Reset: 1754769600
Retry-After: 12          # only on 429

Read X-RateLimit-Remaining and slow down before you hit zero, rather than treating 429 as the signal.

Working within the limits

• Queue, don't spray. Push generation jobs onto a worker queue with bounded concurrency rather than firing them from request handlers.
• Back off with jitter. Retrying every client at exactly Retry-After re-synchronises the stampede. Add randomness — see errors.
• Separate keys per workload. A batch job on its own key with its own limit cannot starve interactive traffic.
• Batch at the prompt level. Ten classifications in one call beats ten calls, and usually costs fewer tokens.

Long-running requests

Text and image calls complete inline; allow a client timeout of at least 120 seconds. Video and speech return immediately and are polled, so they never hold a connection open. Do not set an aggressive timeout on generation endpoints — an abandoned request still consumes upstream capacity and settles against your balance.

Higher limits

Sustained throughput above the defaults is available on higher plans. If you need a specific ceiling, get in touch with your expected requests per minute, average token counts and which models you intend to call.