DocsGuide
Limits
What bounds a call, a stream, a socket and a key, and how a refusal looks.
Every bound is enforced and every refusal is labelled. The numbers here are the platform's today; a change is announced in the changelog.
| What | Bound | Refusal |
|---|---|---|
| requests per key | 1000 per second at the gateway, bursts to 2000 | 429 rate_limited |
| token requests | 20 per 10 seconds per client address | 429 at the identity provider |
| request body | 2 MiB | 413 payload_too_large |
| answer | 4 MiB | 500 internal (a product that can exceed it pages instead) |
| calls in flight on one socket | 64 | rate_limited on the 65th |
| socket frame from the client | 256 KiB | the frame is refused, the socket stays open |
| a stream or a socket | no timeout | none |
Daily allowances
A product may bound what a key does per day (tokens, runs, pages) and says so in its
own section; the answer carries what is left (…_left_today) and the day resets at
00:00 UTC. A key over its allowance gets 429 rate_limited with a retry detail.
Allowances belong to the account, so several keys of one account share them.
Fairness
The per-key rate limit is abuse protection, not a plan. A plan's number is its monthly allowance in units (Usage and units); the console shows an account's usage against it.