OpenAI just put a new sheriff in town, and it works across 40 languages. The company released omni-moderation-latest, a new model built on GPT-4o that upgrades its Moderation API to handle both text and images. It’s free for all developers, with rate limits tied to usage tier. The headline number is a 42% improvement on internal multilingual benchmarks compared to the previous model. For low-resource languages like Khmer and Swati, that figure jumps to 70%. The most dramatic gains were in Telugu, Bengali, and Marathi — languages where the old model barely functioned — showing improvement multiples of 6.4x, 5.6x, and 4.6x respectively. English performance wasn’t left behind, either. The company says the new model’s Spanish, German, and Chinese chops now exceed the old model’s English scores.
What makes this release genuinely useful, not just impressive on paper, is the calibration work. Previous versions gave developers a score, but interpreting what that number actually meant was a guessing game. This model’s probability scores are designed to directly reflect the likelihood that content violates a policy. That consistency will hold across future models too, which is a quiet but critical promise for any team building moderation pipelines. No one wants to retune their thresholds every time an API updates.
The multimodal capability is the other shoe dropping. The model can now evaluate images in isolation or paired with text for violence, self-harm, and sexual content categories. It won’t flag sexual content involving minors — that’s a separate, more sensitive classification pipeline. Two new text-only categories are also live: “illicit” for things like shoplifting instructions, and “illicit/violent” for violent wrongdoing. Companies like Grammarly and ElevenLabs are already named as users, weaving the API into their safety guardrails to catch toxic outputs before they reach users.
OpenAI’s framing here is pragmatic, not utopian. They’re not pretending automated moderation replaces human judgment. The pitch is simpler: these systems help enforce platform policies and reduce the psychological toll on human moderators. That’s the real value proposition — not perfect detection, but sustainable triage. The model is available now through the standard Moderation API endpoint. If it delivers on its multilingual promises, this quietly becomes one of the more consequential free tools in the developer ecosystem.
