GPT OSS Safeguard 20B is a specialized open-weight reasoning model from OpenAI designed for safety classification tasks, released as a research preview alongside a larger 120B variant. It is a fine-tuned derivative of OpenAI's broader gpt-oss open model family and is distributed under the permissive Apache 2.0 license, with both parameter sizes downloadable directly from Hugging Face. This lineage positions it as an accessible, customizable option for teams that need to incorporate nuanced safety logic into their applications without training bespoke classifiers from scratch.

What distinguishes GPT OSS Safeguard 20B is its policy-as-input design: rather than baking a fixed safety rubric into its weights, the model interprets a developer-supplied policy at inference time and classifies user messages, completions, or entire conversations against that policy. It uses chain-of-thought reasoning to arrive at each decision, and that reasoning trace is exposed to developers for review, making it easier to audit and debug classification outcomes. Because the policy lives outside the model, developers can iterate on their rules and refine coverage without retraining, which is well suited to evolving content moderation, compliance, and trust-and-safety pipelines where requirements shift frequently.