Errors

Errors use the OpenAI format, with two additions: retryable and errors.

{
  "error": {
    "message": "Unsupported value for duration. Allowed: 4, 5, 6, 8",
    "type": "invalid_request_error",
    "param": "duration",
    "code": "unsupported_value",
    "retryable": false
  }
}

Field	Description
`message`	Human-readable, in the request’s `Accept-Language` (English or Chinese). Don’t parse it
`type`	Error category, as in OpenAI
`param`	The parameter at fault, as a path such as `messages[0].content`
`code`	Stable error code. Branch on this
`retryable`	Whether retrying the same request may succeed
`errors`	Every validation error, when there are several

Every response carries x-request-id; include it when you contact support.

Status codes

Status	`type`	Common codes
400	`invalid_request_error`	`invalid_parameter`, `missing_parameter`, `unsupported_parameter`, `invalid_json`, `unsupported_media_type`, `unsupported_value`, `input_not_supported`, `too_many_inputs`, `conflicting_inputs`, `invalid_image_count`, `image_count_too_large`, `content_policy_violation`, `input_unreachable` (the model could not download a media URL in the request), `input_too_long` (speech text over the model’s limit), `video_clone_source_invalid` (the reference video can’t be used), `upstream_rejected_request`, `invalid_webhook_url`, `webhook_endpoint_limit_reached`, `project_archived`, `cannot_delete_current_key`
401	`authentication_error`	`missing_api_key`, `invalid_api_key`
402	`insufficient_quota`	`insufficient_quota` (balance too low for the hold), `api_key_limit_reached`, `workflow_budget_exceeded` (the workflow segment would exceed `budget_usd`)
403	`permission_error`	`permission_denied`, `request_blocked`
404	`not_found_error`	`model_not_found`, `generation_not_found`, `file_not_found`, `project_not_found`, `api_key_not_found`, `webhook_endpoint_not_found`, `workflow_not_found`, `workflow_run_not_found`, `workflow_step_not_found`, `drama_not_found`, `drama_episode_not_found`, `drama_artifact_not_found`, `drama_media_not_found`, `video_clone_not_found`, `video_clone_revision_not_found`, `video_clone_asset_not_found`, `unknown_endpoint`
409	`conflict_error`	`generation_not_cancellable`, `workflow_invalid_state` (the workflow run is not in a state that allows the operation, for example confirming a run that is not `waiting_input`), `email_already_registered`
413	`invalid_request_error`	`file_too_large`
429	`rate_limit_error`	`rate_limit_exceeded` (too many requests this minute), `concurrency_limit_exceeded` (too many requests in progress)
500	`server_error`	`internal_error`
503	`service_unavailable`	`model_overloaded` (no capacity available right now), `model_not_available` (model can’t be called yet)

Rate limits

Limits apply per organization and depend on its plan. They cover the inference endpoints: POST /v1/chat/completions, /v1/images/generations, /v1/images/edits and /v1/videos/generations.

Requests per minute. Every inference request returns the same headers as OpenAI:

Header	Description
`x-ratelimit-limit-requests`	Requests allowed per minute
`x-ratelimit-remaining-requests`	Requests left in the current minute
`x-ratelimit-reset-requests`	Time until the count resets, such as `12s` or `250ms`

Over the limit, the request fails with 429 rate_limit_exceeded, and remaining is 0.

Requests in progress. Each accepted request, including an asynchronous job, holds a slot until it finishes. When all slots are taken, new requests fail with 429 concurrency_limit_exceeded.

Both 429 responses carry a retry-after header in seconds. Rejected requests are not charged.

Retrying

Retry only when retryable is true, typically 429, 500 and 503, with exponential backoff. For 429, wait at least retry-after seconds.
Don’t retry other 4xx errors without changing the request.
For asynchronous requests, send an Idempotency-Key so a retry never creates a second job.

Failed generations

An asynchronous job that fails after being accepted has status: "failed" and an error_code such as content_policy_violation, input_unreachable, upstream_rejected_request, model_overloaded or generation_timeout. Failed jobs are not charged.

Failed workflow steps

A workflow step that stops or fails has an error_code on the step, and a failed run repeats it in the run’s error_code.

Code	Meaning
`workflow_charge_exhausted`	The segment used up its hold. The step waits in `waiting_input`; confirm to continue
`workflow_budget_exceeded`, `insufficient_quota`, `api_key_limit_reached`	Not enough budget, balance or key limit. The step waits in `waiting_input`
`workflow_step_failed`	The step failed
`drama_invalid_output`, `video_clone_invalid_output`	The model’s output did not match the required format
`drama_media_failed`, `video_clone_media_failed`	Some images or videos failed to generate
`drama_compose_failed`, `video_clone_compose_failed`	The final video could not be composed
`video_clone_source_invalid`	The reference video could not be imported or read

A step can also carry the code of the API error that stopped it, such as content_policy_violation. Steps that fail with a retryable error are retried automatically before they are marked failed.

Get started

Generate

Workflows

Async and webhooks

Account

Status codes

Rate limits

Retrying

Failed generations

Failed workflow steps

​Status codes

​Rate limits

​Retrying

​Failed generations

​Failed workflow steps

Status codes

Rate limits

Retrying

Failed generations

Failed workflow steps