Errors

Errors use the OpenAI structure, so existing SDK error handling works unchanged:

{
  "error": {
    "message": "model not found: whisper-small",
    "type": "invalid_request_error",
    "param": "model",
    "code": "model_not_found"
  }
}

Status codes

Status	Code	When
`400`	`invalid_request`	Malformed JSON or an unsupported parameter.
`401`	`unauthorized`	Missing or invalid API key.
`404`	`model_not_found`	The requested model does not exist.
`408`	`request_timeout`	The request exceeded the server's time budget.
`429`	`rate_limit_exceeded`	Request or token quota exceeded. Includes `Retry-After` and `x-ratelimit-*` headers.
`429`	`capacity_exceeded`	The node is at its concurrency limit. Includes `Retry-After: 1`.
`502`	`runtime_error`	Inference failed after validation.
`503`	`not_ready`	The service is starting or a required model failed to load. Retry shortly.

Unsupported parameters are rejected explicitly with invalid_request rather than silently ignored: if a request succeeds, every field in it was honored.

Retries

On 429, wait for Retry-After seconds before retrying. On 503, retry with backoff; the service reports ready as soon as its models are loaded.