Skip to main content
This applies to the GPT Image family: gpt-image-2.6-sunburst / gpt-image-2.6-flare, gpt-image-2.5-sunburst / gpt-image-2.5-flare, gpt-image-2, and reverse-engineered routes such as gpt-image-2.6-all. Official-relay routes (2.6 / 2.5 / 2) are moderated by OpenAI; reverse-engineered routes are moderated by their respective providers (-all runs on the ChatGPT web line, -vip runs on the Adobe Firefly line). Status codes and strictness vary by route, but none of these decisions come from APIYI.

Short answer

  • Moderation comes from the provider, not from APIYI’s own rules. Official-relay GPT Image (2.6 / 2.5 / 2) is judged by OpenAI; reverse-engineered -all / -vip routes are judged by their respective providers. APIYI relays the request and passes provider errors through; it cannot overturn a decision or whitelist a prompt.
  • It cannot be turned off. OpenAI provides no switch for official-relay routes, reverse-engineered routes are still bound by their providers’ policies, and APIYI cannot disable moderation on your behalf. moderation only accepts auto / low; low lowers strictness but is not a bypass.
  • Slow failures usually come from output-side moderation: the image is generated first and filtered afterward, so a failed request takes about as long as a successful one. The gateway now returns 400 / 451 as soon as a safety rejection is confirmed instead of retrying until timeout.

Upstream makes the call; APIYI relays the result

Official-relay GPT Image content safety is enforced by OpenAI on the provider side, at two stages: Reverse-engineered routes are not OpenAI official relay, but they also have moderation from their providers: gpt-image-2.6-all runs on the ChatGPT web line, and gpt-image-2.5-sunburst-vip runs on the Adobe Firefly line. The decision maker and strictness can differ. APIYI relays the request to the provider and passes the provider’s response and error back unchanged. There is no “ignore moderation” switch on the platform for a request that the provider has rejected; there is no off switch in the parameters either, and changing groups does not disable moderation.
The platform-level content safety policy is described in Content safety. This page covers the provider’s decision behind GPT Image’s 400 / 451, not a custom APIYI threshold.

Why it is slow and then fails

Output-side review explains the latency

Output-side review runs after the image is generated. The model has already produced the image when the classifier decides whether it can be returned. A failed request therefore spends the same generation time as a successful one, which is why you may wait a long time only to receive an error. In existing tests, 451 image_safety failures took 80–95 seconds, and some complex reference-image cases exceeded 200 seconds. Successful requests can fall in the same range, so do not use latency alone to identify moderation; read the error message and status code first.

Retries add to the total wait

When the provider returns a retryable error, the gateway performs a limited retry; some reverse-engineered routes also retry on another account, which adds more latency. The current optimization is to return 400 / 451 as soon as a content-safety rejection is confirmed, instead of letting the request run until timeout. If you still see a long wait before the error, the time is mainly generation plus output-side review, not an infinite retry loop.

Common 400 and 451 responses for content safety

The status code may differ, but the handling is the same: do not retry the same input; change the prompt or reference image first.

Can it be turned off or lowered?

It cannot be turned off. OpenAI provides no switch for official-relay routes, reverse-engineered routes are still bound by their providers’ policies, and APIYI cannot bypass the provider’s decision. The only knob is moderation, and its effect is limited. On the GPT Image text-to-image API, it accepts only two values: moderation: "low" lowers filtering strictness but does not guarantee a pass; it is especially unreliable for output-side blocks. In existing tests, adding it still returned 400 safety_violations=[sexual]. Do not rely on it in production.
If your requirement is to completely disable content moderation, the GPT Image family cannot meet it. This is not a parameter problem; changing tokens, groups, or routes cannot turn moderation off. You would need to evaluate another model or deployment option.

What to do with a 400 / 451

1

Confirm it is content safety

Check the status code and error.code / message. If it contains safety_violations, moderation_blocked, content_policy_violation, or image_safety, it is a content-safety block, not a parameter or network error.
2

Do not retry the same input

The same prompt and reference image will most likely produce the same result. Output-side review has some randomness, but blind retries only add wait time and cost.
3

Describe the image instead of deleting words

For people, add explicit clothing, age, and scene details. Avoid relying on negative phrases such as “not nude” or “not sexual.” Change one variable at a time; see the linked guides for examples.
4

Regress 2 to 3 times

Output-side review looks at each generated image, so one pass does not prove stability. If the same input fails repeatedly, stop resubmitting it unchanged.
Tier the client timeout by quality so a slow generation is not misread as a failure: at least 120 seconds for low, 240 seconds for medium, and 600 seconds for high. See GPT-Image series overview for the full table.

FAQ

The web app is an agent: a chat model rewrites and expands your prompt first, often adding clothing and scene details. The API sends your original text directly to the image model. The two routes do not send the same prompt, so “the web app works” does not mean “the API is stricter.” See Safety rejection troubleshooting.
No. Those routes also have content-safety policies from their providers; only the provider and strictness differ. A different model may produce a different result, but it is not a moderation bypass and cannot guarantee that the same content will pass.
It only lowers filtering strictness; it is not a bypass. Output-side review evaluates the generated image, and the parameter has no reliable effect on that classifier. Existing tests still returned 400 safety_violations with low.
Output-side review examines a newly generated image each time, so different visual details can lead to different decisions. This is not random approval by the platform, and retrying is not guaranteed to pass. If it fails repeatedly, adjust the visual description instead of relying on luck.
Send the Request ID from the error (or the X-Shellapi-Request-Id response header) and the request body to support; we can check the route and the source of the error. However, the platform cannot promise to override the provider’s decision, and there is no per-prompt whitelist. See How to find the Request ID.

Can an image be rejected after generation?

How 451 image_safety output-side review works, with an anonymized retest and clothing-prompt fixes.

Safety rejection troubleshooting: find the trigger word

Identify the review stage and locate the trigger word with 20 controlled calls.

GPT-Image series overview

Model specs, error codes, and client timeout tiers.

Content safety policy

APIYI’s content safety and compliance requirements.