OpenAI announced on 26 September 2024 that it is upgrading its Moderation API with a new multimodal moderation model, omni-moderation-2024-09-26, which can process both text and image inputs [source]. The company stated the new model achieves higher accuracy than its predecessor, omni-moderation-latest, across eleven policy categories including hate, harassment, self-harm, sexual content, and violence.

According to OpenAI, the updated model reduces false positives and false negatives compared to the previous version. The company reported that omni-moderation-2024-09-26 shows improved performance on internal benchmarks, though specific error rate figures were not disclosed in the announcement.

The Moderation API is designed to help developers identify content that violates OpenAI's usage policies. The multimodal capability allows the model to evaluate images alongside text, which OpenAI described as useful for applications processing user-generated visual content.

OpenAI stated that the new model is available immediately through the Moderation API at no cost to developers. The company noted that the previous model, omni-moderation-latest, will continue to be supported but recommended migration to the new version for improved accuracy.

The announcement did not specify whether the model upgrade was prompted by observed failures or accuracy issues with the prior version. OpenAI indicated that the model was trained on updated data and benefits from architectural improvements, but did not provide technical details about the training process or the nature of the improvements.

Developers using the Moderation API can access the new model by specifying omni-moderation-2024-09-26 in their API requests.