GPT Image moderation_blocked vs Sume content_policy_rejected

When OpenAI blocks a GPT Image request it returns error.code = "moderation_blocked", sometimes with error.moderation_details. Do not retry unchanged: edit the prompt or input images. On Sume, the matching refusal is grouped under the public reason content_policy_rejected, and the same advice applies.

The OpenAI side is from its image generation guide; Sume's from Image generation and the OpenAPI schema, read 2026-10-01.

What does OpenAI's error contain?

The guide shows error.type of image_generation_user_error, error.code of moderation_blocked, and an optional moderation_details object with a moderation_stage and coarse categories such as harassment. The stage is input when the prompt or request inputs were blocked, output when a generated image was, and unknown as a rare fallback. OpenAI says this gives coarse debugging context without internal classifier labels or scores, and recommends a generic end-user message.

What does Sume return?

Sume's OpenAPI describes a public_reason field as machine-friendly and groupable, with examples generation_rejected, image_content_rejected and content_policy_rejected. The job error carries content_policy_rejected when a provider rejection is classified as a content-policy refusal. I found no Sume field that mirrors moderation_stage or categories, so do not expect the input-versus-output split.

How do the two compare?

What should my code do?

Branch on the public reason, not on message text. Show the user a generic message, log the request id and reason for support, and ask for a changed prompt or image before resubmitting. Sume's image docs state that failed generations are not billed, so a refused request does not leave a charge. The moderation setting itself is covered in GPT Image 2.5 moderation on Sume.