Rate Limits

Per address

Flow AI accepts up to 500 requests per second from one IP address, with short bursts up to 1,000. Past that, requests are refused with 429 and the code rate_limited until the rate drops.

Per key

Keys have no request-rate limit of their own. What a key can do is bounded by its spending limit and your balance.

Model hosts

Each host has its own capacity. When a host is busy, Flow AI tries another host for the same model where there is one; if none can answer, you get 503 with the code host_unavailable. Retry with backoff.

Timeouts

A request can run for up to 10 minutes. For long answers, use streaming so text arrives as it’s written.