OpenAI Moderation API: Free, Fast, and the First Guardrail Layer You Should Turn On

If your AI application still doesn't have any content guardrail, OpenAI's Moderation API is the cheapest possible starting point: it's free, and there's almost no excuse not to add it.

Free, and it already covers text and image

OpenAI's Moderation API remains free in 2026, and now covers both text and images through the omni-moderation model. It's a free classification tool, accessed through a dedicated moderation endpoint on the OpenAI API, that evaluates whether text (and, for the newer models, images) is potentially harmful across a set of content categories.

The categories it classifies

The API classifies text content into categories that include hate, harassment, self-harm, sexual content, and violence — returning both boolean flags and per-category confidence scores, which gives you the flexibility to decide where to set the blocking threshold yourself instead of accepting a fixed binary decision from the provider.

Latency makes it viable on the critical path

The Moderation API is fast, with typical latency of 15-25ms, adding negligible overhead when run synchronously on user input — unlike heavier guardrails that require an extra call to a full LLM, this can run on the critical path of every request without noticeably degrading the user experience.

The fine print: not everything works the same for image and text

Image moderation is supported, but some categories are text-only — meaning not every content category works the same way for text and image inputs. It's worth checking the specific documentation on which categories apply to each input type before assuming full coverage for both.

Why it's the “minimum common denominator” guardrail

The Moderation API doesn't replace more sophisticated guardrails like Bedrock Guardrails or Azure's Prompt Shields — it doesn't specifically detect prompt injection or jailbreaks. But as a free baseline layer against clearly harmful content (hate, violence, explicit sexual content), there's no real reason not to have it running on any application that exposes an LLM to the public, even as a complement to more advanced guardrails rather than a substitute for them.