Errors

Errors follow OpenAI's shape: an error object with message, type and code.

Status	Code	What to do
400		Fix the request.
401	`invalid_api_key, key_expired`	Check the key.
402	`insufficient_credit`	Add credit, or turn on auto-reload. Lowering max_tokens reserves less.
402	`spend_limit_reached`	Raise or clear the monthly spend limit on the account, project or key.
403	`admin_key_required, inference_key_required`	Use the other kind of key.
403	`policy_blocked`	One of your enforcing policies matched. policy_id says which.
403	`platform_policy_violation`	The platform rule matched. It can't be turned off.
404	`model_not_found`	See GET /v1/models.
413	`request_too_large`	Send a smaller request. The limit is 8 MB.
429	`rate_limit_exceeded`	Wait for Retry-After seconds.
503	`model_not_deployed`	This model isn't running on the platform at the moment.
503	`no_capacity`	No GPU was free. You weren't charged. Retry after Retry-After seconds.
503	`gateway_unavailable`	The API is starting up or briefly down. Retry after Retry-After seconds.
503	`policy_evaluation_failed`	An enforcing check couldn't finish, so the request stopped. Retry.
504		The request hit the time limit, and is billed up to it.

If a stream fails partway, its last event before [DONE] is an error object, which the OpenAI SDKs raise as an exception.