Guides
Rate Limits
What happens under sustained load, and how to handle it.
What's enforced today
Prismma does not enforce a hard, published per-key requests-per-minute or tokens-per-minute limit today. Your requests are served from a shared model-provider pool with finite throughput — under sustained, high-volume traffic you may occasionally see a 429 if the pool is momentarily saturated, independent of your own usage.
This may change as usage grows — treat 429 as a signal to retry, not as a hard quota you've exhausted (that's a 402, see Errors).
Handling 429
- The official OpenAI and Anthropic SDKs retry 429/5xx with exponential backoff automatically — you don't need to add your own retry loop for typical usage.
- For production workloads issuing many concurrent requests, add jitter to your own retry logic to avoid synchronized retry storms.
429 (rate limited) and 402 (out of credit) are different problems — check the HTTP status and the response body's code field (Errors) before treating a 429 as a billing issue.