GPT OSS Safeguard 20B is a safety reasoning model built as a fine-tuned variant of the gpt-oss open-weight base. Rather than baking a fixed set of safety rules into the model, it reads a developer-provided policy at inference time and uses chain-of-thought reasoning to classify content—including user messages, completions, and full conversations—according to the policy you supply. This means the same model can enforce very different safety standards depending on the policy it receives, allowing developers to draw the policy lines that match their specific use case and iteratively refine them without retraining. The model exposes its reasoning steps in full, so teams can review how decisions are reached, and it supports adjustable reasoning effort levels (low, medium, high) as well as Structured Outputs for easier integration into production pipelines.

The model is post-trained from the underlying gpt-oss checkpoints and was developed with feedback from the open-source community, released under a permissive Apache 2.0 license that grants broad rights to use, modify, and deploy. It was trained to operate within the harmony response format and is compatible with the Responses API, though it is not intended to serve as a general chatbot—its sweet spot is safety classification pipelines, content moderation, and policy enforcement layers where the original gpt-oss base model would handle end-user interaction. Developers receive guidance on crafting policies through OpenAI's prompting cookbook, and the flexibility of swapping policies at inference makes this approach more adaptable than training a static classifier from large labeled datasets. With 21B total parameters and 3.6B active parameters during inference, the 20B size offers a lower-latency option compared to its 120B sibling for teams that need safety reasoning without the heavier compute footprint.