Skip to main content

​
Requests-per-minute limits

Primary rate-limiting is a 60-second sliding window in Redis. Two halves worth keeping apart:
  • the limit comes from the plan and from any quota preset attached to the API key you are calling with;
  • the counter is per workspace, shared by every key in it.
So a second key raises nothing on its own — if you need budgets that cannot starve each other, use separate workspaces. Limit varies by plan and any attached quota preset: You see the effective value in API Keys next to each key.

​
Exceeded response

403, not 429, and there is no Retry-After header. Client libraries that back off on 429 — the OpenAI SDK included — treat this as a hard client error and do not retry it at all. Branch on error.code, not on the status: several unrelated refusals share this 403.
A refused request is counted into the window too, so retrying inside those 60 seconds occupies another slot and pushes your recovery further out. Wait out the window, or stay under the limit — see Rate limits & retries for the playbook.

​
Daily token budgets

In addition to requests per minute, some plans cap total tokens per day (rateLimitTpd in your quota preset). Hitting this is a 403 with its own code, and the counter resets at UTC midnight:

​
Global safety net

We enforce a 5 000 req / 10 min per IP ceiling at the edge to stop scraping. This almost never trips for real users — only poorly-configured crawlers.

​
Best practices

  1. Don’t retry inside the window. There is no Retry-After to honour, and a retry sooner than 60 seconds extends the wait rather than shortening it.
  2. Queue on your side. A local token bucket at ~80% of your limit is the only approach that beats the window instead of feeding it.
  3. Use exponential backoff with jitter for 5xx — 1s → 2s → 4s → 8s (+random 0–500 ms) — but only where repeating the call is safe. A 500 on a generation call may already have been billed; see Rate limits & retries.
  4. Consider fallback chains. They cover a provider’s refusal on gpt-4o by advancing to gpt-4o-mini or gemini-2.5-flash. They do not cover your own workspace limit, which is refused before a model is chosen.
  5. Separate environments by workspace, not just by key. Keys share their workspace’s window; only a separate workspace gives dev its own.

​
Rate limits on specific endpoints

  • /contact/sales — 5 req / hour per IP (spam protection)
  • /public/plans and /public/models — 30 req / min per IP
  • /v1/files upload — the RPM limit applies; additionally serialised by workspace (one upload at a time), so a concurrent upload answers 409 upload_in_progress — that one is safe to retry in a moment, and our SDKs do