Rate Limits
How many requests the API accepts, and what to do when you reach a limit.
These are the current beta limits. They may change as we learn how the API is used.
| Limit | Requests per minute |
|---|---|
| Per IP address | 3,000 |
| Per organization | 1,200 |
| Per API key or OAuth connection | 600 |
Every limit applies at the same time, and the REST API and the MCP server share them. More keys or connections don't add to your organization's limit. A request refused with 401 or 403 doesn't count against the organization, key and connection limits.
When you reach a limit
The API answers 429 rate_limited. The error's data.retryAfterSeconds says how many seconds to
wait, and the Retry-After header says the same. Wait that many seconds, then retry. This wait and
the error's detail are the authoritative answer, even if the numbers above change.
On the MCP server, a tool call over the organization, key or connection limit gets the error as the
tool's result. A tool result has no header, so the wait is in data.retryAfterSeconds, and the
error's detail states it for the assistant (see MCP Server).
To stay below the limits:
- Spread requests out instead of sending them in bursts.
- Use
limit=500to fetch more per request. - Re-read only the window you need (see Data and freshness).
Last updated on