Rate limits

The API takes a given number of calls in a given time. A budget refills steadily, not all at once: a refused call only waits for the next one to come back, and the answer says when.

NameBudget
calls300 requests per 5 minutes per accountRefills steadily: one back every second.
writes60 write and destructive operations per hour per accountRefills steadily: one back every minute.
by_ip1200 requests per 5 minutes per IP addressRefills steadily: four back every second.
RateLimit-Policy
The budgets that apply to the request: size, window and name.
RateLimit
What is left of a budget, and in how many seconds it is full again.
X-RateLimit-Limit
How many calls that budget takes at once when it is full, as older clients read it.
X-RateLimit-Remaining
What is left of it.
X-RateLimit-Reset
In how many seconds it is full again.
Retry-After
With a 429: in how many seconds the next call passes.

REST: the request is refused with 429; Retry-After says in how many seconds the next call passes.

MCP: the tool returns an error (isError: true) with a text the model can read, so an agent can wait and try again instead of giving up.