Reference
Rate limits
What is limited, and what happens at the edge.
| Surface | Limit | Window |
|---|---|---|
| Gateway inference | 600 requests | per minute per key |
| Control plane reads | 1,000 requests | per minute per tenant |
| Key minting | 60 requests | per minute per tenant |
A limited request answers 429 with a retry-after header. The limit is on the reservation substrate rather than on a counter beside it, so a burst cannot slip through between the check and the hold.
http