Rate limits and quotas
Every request is admitted against two ceilings — your key's own, and your account's allowance on the model — and the lower one applies.
Per key
| Limit | Default | On exceeding |
|---|---|---|
| Requests per minute | 1,250 | 429 rpm_exceeded, with Retry-After. |
| Tokens per minute | 5,000,000 | 429 tpm_exceeded. Counted on the estimated prompt at admission, not after the fact. |
| Images per minute | 60 | 429 ipm_exceeded. Image generation only, counted in images rather than tokens. |
| Active keys per account | 50 | 409 when creating the fifty-first. |
All three ceilings are token buckets that refill continuously, so a burst is absorbed rather than refused at a minute boundary. A key may be given stricter limits than these — worth doing for one handed to a batch job — while raising one above the default is a conversation with support, because the ceiling is our side of a bargain with the model vendor rather than a preference. Set them per key from the keys page, which also shows what each model allows you.
Per model
Each model also carries an allowance for your whole account, counted per model. It is there because the capacity behind a model is finite and shared, and on several models it is the lower of the two ceilings — so a key showing a million tokens a minute can still be refused on a model whose allowance is smaller. Exceeding it is a 429 with model_rpm_exceeded, model_tpm_exceeded or model_ipm_exceeded and a Retry-After; the message names the figure.
Per request
| Limit | Value |
|---|---|
| Request body | 50 MB |
| Attachments | Counted by what they are, not by the size of the file — see below |
| Context window | The model's own — up to 1,048,576 tokens |
| Max output | The model's own — up to 393,216 tokens |
| Completions (n) | 1 — above it is a 400, not a cap |
| Time to first byte | Up to 10 minutes; reasoning models are expected to think |
max_tokens above a model’s limit is clamped rather than rejected, and the clamped value is what is sent upstream. For a request that switches thinking off on a model that honours that, it makes the reservation a bound on the bill. Otherwise it is a sized estimate: reasoning is not bounded by max_tokens, so the reservation adds a measured allowance for it and an unusually long reasoning run can settle slightly above what was held. The models that reason whatever they are asked — GLM 5.3 Flash — are reserved that way even with thinking off, because there max_tokens bounds nothing. See models and pricing.Credit, not quota
There is no monthly quota to exhaust. Spending is bounded by your balance: a request whose worst case exceeds it is refused with a 402 before it is sent. That worst case is held for the duration of the request, so concurrent calls cannot each spend the same credit. A balance can still end slightly below zero when a reasoning run settles above the allowance held for it; the shortfall comes off the next top-up rather than being written off, and nothing is cut off mid-stream to prevent it. Turn on automatic top-ups if you would rather not watch the balance, with a daily ceiling you set.
Need more?
The defaults are conservative because they are the ones a new account gets without asking. If you are planning something larger, tell us the shape of it and we will raise them — there is no tier to buy.
Last updated 2026-10-09.