Source: OpenAI moderation guide

OpenAI's guide to the free omni-moderation endpoint: 13 categories, per-category scores, image support and inline moderation in the Responses API.

Contents

Summary

The moderation endpoint classifies text and images against a fixed harm taxonomy and returns a boolean flagged, per-category booleans and per-category scores. It is free. It can be called standalone or attached inline to a Responses API request.

Key claims

  • Model omni-moderation-latest takes text and images (images up to 20 MB); it does not classify audio.
  • 13 categories: harassment, harassment/threatening, hate, hate/threatening, illicit, illicit/violent, self-harm, self-harm/intent, self-harm/instructions, sexual, sexual/minors, violence, violence/graphic.
  • Only self-harm, sexual and violence categories score images; others are text-only and score 0 for image-only input.
  • Response fields: flagged, categories, category_scores (0 to 1), category_applied_input_types.
  • Scores should feed your own policy rather than auto-block; the model is upgraded continuously, so custom thresholds may need recalibration.
  • Inline moderation covers tool-call arguments and tool outputs but not tool names, descriptions or schemas; with streaming, results arrive only after the full output; a refusal that discusses harmful content can still be flagged.
  • Not designed for CSAM detection.

Feeds

Content moderation · Guardrail tooling

Suggest changes →