Rate limits

Only endpoints that intentionally differ from the standard API carry a chip; the chip links to the note that explains the difference.

Every budget below renews on a fixed window and answers 429 when spent. Find your number in the table, check whether that 429 carries a Retry-After header, and react accordingly. A 402 is never a rate limit — it means your balance is empty, so top up instead of retrying.

Your budgets at a glance

Your inference budget belongs to your account, not to any single key: every request counts against the owning user's per-minute window, so minting extra keys never multiplies it. Omit rate_limit when creating a key and it resolves to 500; an explicit value must be between 1 and 1000 — anything else (including 0 and negatives) is rejected with 400. The catalogue read (GET /v1/models) carries no per-minute budget.

What a 429 looks like

Inference refusals quote your account's own limit back at you. On the OpenAI surface:

{
  "error": {
    "message": "rate limit exceeded: 500 requests/min for your account (all keys combined) — wait a minute or contact support to raise it",
    "type": "rate_limit_exceeded"
  }
}

The Anthropic surface carries the same message in its own envelope:

{
  "type": "error",
  "error": {
    "type": "rate_limit_error",
    "message": "rate limit exceeded: 500 requests/min for your account (all keys combined) — wait a minute or contact support to raise it"
  }
}

Playground refusals are shorter (rate limit exceeded with type rate_limit_exceeded), and tired-balance refusals say too many concurrent requests for current balance — wait for in-flight requests to complete or top up. Free-model refusals carry the promo message with type rate_limit_exceeded and code free_tier_quota_exceeded on the OpenAI surface (rate_limit_error on Anthropic). Signup refusals use code rate_limited.

Staying under

Honor Retry-After wherever it is present; otherwise back off exponentially — the full retry pattern lives under Errors and is not repeated here. Keep concurrent requests serialized while your balance is thin, or top up past $0.25 to lift the concurrency cap entirely. Paid models are never gated by the free-model quota: a free-model 429 names its own remedy in the message.

Where these numbers come from

These are the live gateway defaults; if one changes, this page changes with it.