Limits protect each cell from a burst, yours or another customer’s. They are
per key: two keys from the same workspace do not share a quota. When you
exceed a request limit, the API responds 429 rate_limited with a Retry-After header that says
how many seconds to wait.
| Limit | Value | Scope | When exceeded |
|---|---|---|---|
| API requests | 120 per minute | Per key | 429 rate_limited, Retry-After |
Requests rejected with 401 |
60 per minute | Per IP address | 429 rate_limited, Retry-After |
| Time to open a realtime session | 60 seconds from creation | Per session | realtime_session_invalid on connect |
| Duration of a realtime session | 900 seconds | Per session | The server hangs up with session.ended, reason: max_duration |
API requests
Each key has its own budget of 120 requests per minute. Two keys behind
the same IP do not share a quota. A request rejected with a 4xx counts too.
Requests rejected with 401
Requests without a key, with a malformed key or with a key from another product are counted separately, per IP address, with a limit of 60 per minute. That way a misconfigured client cannot exhaust the budget of the valid keys coming from the same network.
Realtime sessions
POST /agents/v1/realtime/sessions returns a url that you must open within 60 seconds. A
conversation lasts at most 900 seconds: after that time, the server sends
session.ended with reason: max_duration and closes the socket.