OpenAI Moderation API

Send text or an image to the moderation endpoint and it returns a flag per category (hate, harassment, self-harm, sexual, violence, and sub types) with a score for each. The current model is multimodal, so an image and its caption can be judged together. It's free, it's fast, and it's the same classifier OpenAI uses to police its own API traffic.

The catch is that the categories are fixed. You can't teach it your policy, you can only threshold what it gives you. It's excellent as a cheap first filter in front of a chatbot or a comment box and as a way to keep your own LLM app inside OpenAI's usage rules.