Errors

Status codes, what to retry and what not to.

Errors use the OpenAI error shape, so an existing client's handling works unchanged. A provider's own error body is forwarded verbatim; everything we generate ourselves is a fixed message with a stable code.

Status codes

Field	Type	Description
`400`	invalid_request	Malformed body, or a parameter the model does not support.
`401`	unauthorized	Missing, malformed, revoked or expired key.
`402`	insufficient_credits	The wallet or the key's spend limit cannot cover the request.
`403`	forbidden	The model is disabled, or not on this key's allow-list.
`404`	not_found	No such model or endpoint.
`429`	rate_limited	Key rate or concurrency limit. Honour Retry-After.
`5xx`	upstream_error	The provider failed. The hold is released in full.

Retrying safely

Retry 429 and 5xx with exponential backoff and jitter.
Never retry 400, 402 or 403. They will not succeed on a second attempt.
A failed request settles at zero. Retrying does not double-charge you for the attempt that failed.

Request ids

Every response carries an X-Request-Id. Quote it when asking about a specific call; it is the key to the whole log line, including the model actually served and the exact amount charged.