Errors

Error format

Every error response carries a JSON body in this shape:

{
  "error": {
    "code": 20002,
    "message": "Missing required parameter",
    "request_id": "req_8a4f0d2e-1c8b-4f1a-9e2d-3b7c5a6f0a1b"
  }
}

The HTTP status code reflects the category (401 / 400 / 402 / 429 / 500 / 502 / 503 / 504); the code field gives you the precise reason.

For failed tasks, the same envelope is returned inside the polling response under error — see the model docs' Response section.

Code catalog

1xxxx — Authentication (HTTP 401 / 403)

2xxxx — Validation (HTTP 400)

3xxxx — Billing (HTTP 402 / 400)

4xxxx — Resource (HTTP 404)

5xxxx — Rate limit (HTTP 429)

6xxxx — Server internal (HTTP 500)

7xxxx — Capacity (HTTP 503)

8xxxx — Workflow execution (returned inside the polling response body when status="failed")

The 8xxxx codes are written to the task by the worker; polling returns them under error.code once status="failed".

Handling 80006 content-policy errors

80006 is a cross-model workflow code. It does not, by itself, identify a particular model, provider, or moderation stage as the cause. A safety check may have rejected any of these:

• the prompt;
• a reference image;
• a reference video; or
• the generated image or video.

Do not retry the same payload unchanged. First remove or rephrase sensitive content and verify that every reference image and reference video is suitable. Even when all submitted inputs are acceptable, a generated output can still trigger the code.

For models whose own API documentation lists the boolean content_filter parameter, direct API-key requests can also try the less restrictive route:

{
  "model": "<a model that supports content_filter>",
  "prompt": "<your revised prompt>",
  "content_filter": false
}

This option is model-specific, not a global request parameter. It selects a less restrictive moderation path; it does not disable safety review. The hosted Playground keeps moderation enabled, and upstream policy checks still apply, so a request can still return 80006. Check the selected model's parameter list before sending it.

Refund behavior depends on where the verdict occurred and on the selected model's documented policy. Check the failed task's usage.credits: 0 means the reservation was fully refunded; a positive value means a post-generation charge was retained.

Recommended handling

import requests, time

def safe_call(method, url, **kwargs):
    for attempt in range(3):
        r = requests.request(method, url, **kwargs)
        body = r.json()

        if r.status_code == 200:
            return body

        if r.status_code == 429:
            time.sleep(int(r.headers.get('Retry-After', '5')))
            continue

        if r.status_code in (502, 503, 504):
            time.sleep(2 ** attempt)
            continue

        raise RuntimeError(
            f"{body['error']['code']}: {body['error']['message']}"
        )

    raise RuntimeError("max retries exceeded")

What to retry vs not

Support

When reporting an issue, include:

• The request_id from the response
• The full request URL + method
• Approximate timestamp (UTC)

Email: [email protected]