Skip to main content
Each key has a rate of requests per second, set by its plan in the portal (a key with no rate of its own gets 10). All the keys of one account share the account’s rate: a second key does not double it, and the API, the router and launch count against the same rate. The rate allows a small burst: an account that has been quiet can send two seconds’ worth at once (20 requests on a rate of 10), then one more every tenth of a second. Requests spread evenly at the rate never wait. Above it the API answers 429; the next request is let in as soon as the rate allows, at most a second later. Every request that carries a valid key counts, whatever its answer, a 400 included; a request refused with 429 does not. /, /health and /openapi.yaml are not counted.
There are no rate-limit headers: pace your requests at your rate, and on a 429 wait a moment (a few hundred milliseconds, at most a second) instead of retrying in a loop.

Credits

Usage is paid in credits. Every plan includes credits for each 30-day period, and extra credits can be bought on top of a plan in the portal, which lists what each unit costs. Credits are spent by:
  • on the API, each successful request (any status below 400), the opening of a WebSocket or SSE stream included;
  • each minute a WebSocket or SSE stream stays open;
  • on the router and launch, each request the key lets in, whatever the router answers.
On the API a refused request (a 400, 401 or 429) spends nothing; on the router a request refused for its key or its limit (401, 429) spends nothing. The plan’s credits are spent first; extra credits never expire. When an account has no credits left its keys stop working (401) until the plan renews or extra credits are added.

Streams count once

Opening a WebSocket or an SSE stream is one request. Frames after that do not count against the rate and cost no credits of their own (an open stream is billed by the minute, not by the frame), and one socket carries up to 200 channels. Anything that changes over time is cheaper and fresher as a stream than as a polled REST route:

Batch what you can

  • /tokens?ids= reads up to 50 tokens in one request.
  • /prices?tokens= prices up to 100 tokens in one request.
  • Identical requests from many clients are cheap on the server (it shares one fetch), but each still counts against your key.