OpenAI omni-moderation

omni-moderation is the outlier in OpenAI's catalogue: a model with no price. Released on 26 September 2024, it classifies text and images against a list of harm categories — sexual content, hate, harassment, self-harm, violence and others — and returns a score for each, rather than prose. Developers use it as a gate in front of or behind a generative model, and OpenAI's usage policies effectively assume they will. It replaced the text-moderation family, which had been text-only. Two things changed with it: images became valid input for some categories, and two new text-only harm categories were added, alongside scores the company described as more accurate. The older text-moderation-007, -latest and -stable snapshots were shut down on 27 October 2025, leaving omni-moderation as the only moderation endpoint in the API. Its second life began on 4 June 2026, when OpenAI stopped treating moderation as a separate call. Passing a moderation object in a Responses API or Chat Completions request now returns moderation results for both the prompt and the generated answer in the same response — one round trip instead of three. The model itself is unchanged since the 2024-09-26 snapshot, which for a component that sits in the critical path of other people's safety systems is a feature rather than neglect.

▸Videos

No videos yet.

▸Reviews