#Rate Limits

Capacity can vary by account, group, model, upstream provider, and current availability. There is no single RPM or TPM value that applies to every request.

Use a bounded queue, a concurrency limiter, a total timeout, and a retry budget. Start at low concurrency and increase it while monitoring latency, 429, 5xx, usage, and cost.

A 429 can indicate quota, balance, concurrency, rate, or upstream capacity. Inspect the response and usage logs before assuming it is only an RPM limit.