Usage, rate limits & budgets

Every key carries limits that the gateway checks before a call reaches a model, not after:

Field	Default	Enforced?
`requests_per_minute`	120	Yes, a full bucket returns `429 rate_limit_exceeded`
`monthly_token_budget`	5,000,000 (0 = unlimited)	Yes, a project at or over budget returns `429 budget_exceeded`
`tokens_per_minute`	200,000	Stored on the key, not currently wired into admission, treat it as reserved for a future check, not a live throttle

Both enforced limits are evaluated before the request is sent upstream, so a rejected call never spends a model’s time or a project’s token budget.

Context window

Requests are also checked against a context window before dispatch, by default 4096 tokens, estimated at roughly 4 characters per token. A prompt (or, for threads, the system block plus transcript) that would exceed the window is rejected with 400 context_length_exceeded rather than sent upstream to fail there. See Errors for the exact message shapes.

Scoping limits per key

Override any of these when minting a key (see Authentication & API keys):

{
  "requests_per_minute": 60,
  "monthly_token_budget": 1000000
}

Omit a field to fall back to the gateway-wide default above.