Skip to main content
Rate limits protect the platform from bursts. They are not how spending is controlled — credits do that.

​
The limits

10 requests per second, with a burst of 20. The bucket is per API key, so one key’s traffic cannot exhaust another’s.

​
The headers

Every response carries:
There is no X-RateLimit-Reset. The bucket refills continuously rather than at a fixed boundary, so there is no instant to report.

​
When you exceed it

429, with a Retry-After header in seconds. A rate-limited request is never charged — it did no work. Back off and retry. Retry-After is the honest answer; exponential backoff on top of it is polite.

​
What is not rate limited

Provider-backed work — placing calls, running agents — is governed by your credit balance, not by request count. A rate limit is about protecting shared infrastructure; a balance is about what you have paid for. Conflating the two would mean a well-funded customer being throttled for spending money. Concurrent calls are capped separately, per account, and default to 5.