OpenAI: omni-moderation-latest:free

omni-moderation-latest is OpenAI's next-generation multimodal content moderation model, built upon the GPT-4o architecture. Unlike its text-only predecessors, it supports the joint moderation of text and image inputs. The model offers a 42% improvement in accuracy for non-English languages, excelling particularly in low-resource contexts. It identifies harmful content across 13 categories, including hate, harassment, self-harm, sexual content, violence, and new categories like "illicit" and "illicit/violent." The scores are calibrated (0-1), providing a realistic reflection of the probability that the content violates OpenAI's safety policies.