The GPT OSS Safeguard 120B is a safety reasoning model that represents a shift away from traditional content classifiers toward policy-grounded judgment. Rather than training fixed decision boundaries from labeled examples, this model interprets a developer-supplied safety policy at inference time, applying it to classify user messages, completions, and full conversations. It employs chain-of-thought reasoning that developers can inspect to understand how conclusions are reached, and it supports configurable reasoning effort levels so teams can trade off depth against latency depending on the use case. This design gives developers explicit control over where the policy lines fall, enabling rapid iteration as safety requirements evolve.

The model is a post-trained derivative of the open-weight GPT OSS family, developed with input from the open-source community and released under a permissive Apache 2.0 license. It was initially built for internal use at OpenAI before becoming publicly available, and it pairs with the Responses API for deployment. The 120B variant is optimized for production workloads and can run on a single H100 GPU, making it accessible for teams without massive infrastructure. Unlike models designed for end-user interaction, this model is purpose-built to serve as a safety classification layer behind the scenes, classifying content with justified decisions and supporting structured outputs for integration into broader moderation pipelines.