Upgrading the Moderation API with our new multimodal moderation model

OpenAI just swapped the brains behind its Moderation API for something built on GPT-4o, and the headline change is simple: the thing can now look at pictures, not just parse text. That's a real gap closed. For years developers filtering harmful content had to bolt on separate image classifiers or just hope users didn't upload something nasty in visual form. Now one model handles both.

Accuracy is the other pitch here. OpenAI says the new model catches harmful text and images more reliably than the previous version, which matters a lot once you're running moderation at the scale of a messaging app, a marketplace, or a comment section with millions of posts a day. A few percentage points of improved precision translate into thousands fewer false positives annoying legitimate users, and thousands fewer genuine violations slipping through.

The API itself stays free, which is worth pointing out because content moderation tooling from smaller vendors often comes with per-call pricing that adds up fast for anyone running a community platform. OpenAI positioning this as a foundational, no-cost layer nudges more developers toward using a standardized safety layer instead of rolling their own heuristics or ignoring the problem until it becomes a PR headache.

This isn't a flashy launch — no new chatbot personality, no benchmark chart bragging about reasoning scores. But moderation infrastructure is the unglamorous plumbing that keeps platforms from turning into cesspools, and OpenAI clearly wants to be the default plumbing provider. Given how many apps are now built on top of GPT-4o already, plugging in the same provider's moderation layer is the path of least resistance for a lot of teams.