Reference

Errors

Errors are plain JSON with an error string and an honest status code. Nothing is charged for a request that does not produce a completed generation.

Shape

Any error

{ "error": "model and prompt are required" }

Status codes

Field	Type	Description
`400`	Bad request	Missing `model` or `prompt`, or no provider can serve the model under your routing rules — including a `provider` pin that does not offer it. Fix the request; retrying unchanged will not help.
`401`	Unauthorized	The bearer token is missing, malformed or deleted. See API tokens.
`402`	Payment required	Balance below the quoted cost. The message includes what is needed and what you have; top up and retry the same request.
`404`	Not found	No generation with that id on this account. Ids are scoped to the account that created them.
`502`	Bad gateway	The provider refused the submission — bad parameters for that model, or an upstream outage. Nothing is charged. Retry, or pin a different provider.

Failed generations

A generation that was accepted and later failed is not an HTTP error: the poll returns 200 with "status": "failed" and an errorfield carrying the provider's reason. Handle it in your polling loop rather than in your HTTP error branch.

Retry strategy

Retry 502 and network faults with backoff — they are transient by nature.
Never retry 400 unchanged; the request is wrong, not unlucky.
Treat 402 as a signal to top up, and alert on it: it means production is stopped, not slowed.
Poll timeouts are not failures. A generation that is still pending after your loop gives up is readable later with the same id.

← PreviousBilling