Guides
Rate limits, concurrency, and backoff
Limits can vary by Key, model, group, and upstream state. 51cowork does not publish a fixed numeric promise.
Treat 429 as a temporary capacity signal, honor Retry-After, use jittered bounded backoff, and cap client concurrency.
- Fixed TPS
- Not promised
- Signal
- 429
- Header
- Retry-After
- Policy
- Bounded backoff
Client policy
- Set an account-level concurrency limit and queue.
- Read Retry-After before exponential backoff with jitter.
- Do not blindly retry 400, 401, 402, or 403.
- Retry temporary 5xx only a bounded number of times and record request_id.
Streaming requests
Streaming needs a longer read timeout. First-token wait and long output should not be mistaken for a dead connection.