Errors

Errors use the native shape of the format you called. The Chat Completions, Responses and /v1/models endpoints use OpenAI’s shape:

{ "error": { "message": "Invalid or missing API key.", "type": "invalid_request_error", "code": "invalid_api_key" } }

/v1/messages and /v1/messages/count_tokens use Anthropic’s shape:

{ "type": "error", "error": { "type": "authentication_error", "message": "Invalid or missing API key." } }

Status codes

Status	OpenAI type / code	Anthropic type	When	Retry?
400	invalid_request_error	invalid_request_error	Malformed body, unsupported field value or content type, output limit not a positive integer, body over 32 MB, input over the context window, remote image URL	No: fix the request
401	invalid_request_error / invalid_api_key	authentication_error	Missing, invalid, revoked or expired key, or a disabled account	No
404	invalid_request_error / model_not_found	not_found_error	Unknown model (the message lists the available ones), or an unknown path	No
502	server_error	api_error	The model server failed or is unreachable	Yes, with backoff
503	server_error	n/a	No capacity within 30 s (OpenAI formats). Has Retry-After	Yes, after Retry-After
529	n/a	overloaded_error	No capacity within 30 s (Anthropic format). Has Retry-After	Yes, after Retry-After

A 400 from the model server, such as a prompt over the context window, keeps the model server’s own message when it’s short and safe to show. Otherwise the message is a generic “Request rejected by the model server”.

Errors during a stream

Capacity and validation errors are always returned as normal HTTP errors, before any stream starts. If the model server fails after streaming has begun, the stream ends with an in-band error:

• Chat Completions: the stream stops without data: [DONE].
• Responses: a response.failed event.
• Messages: an error event with api_error.