OpenAI releases gpt-oss-safeguard, open-weight safety reasoning models that read your policy
20b and 120b Apache-2.0 models that classify text against a policy you supply at inference time, with configurable reasoning effort.
For context: this is a late-2025 release that later guard models compare themselves against. The date here, 29 October 2025, is the last-updated date of the Hugging Face collection shown on the model card. The card itself gives no release date, and the repository was created earlier, on 18 September 2025.
gpt-oss-safeguard is a safety reasoning model built on gpt-oss. It classifies text against safety policies you supply at inference time. The card says the model interprets the policy you write, so it can generalise across products with minimal engineering, and it exposes its reasoning. Reasoning effort is configurable as low, medium or high.
There are two sizes. The 20b has 21B total parameters with 3.6B active and fits on 16 GB of VRAM. The 120b has 117B total with 5.1B active. Both are licensed Apache-2.0. The card says the model must be used with the harmony response format or it will not work correctly.
The card limits its scope. It is intended for safety use cases; for other applications it recommends the gpt-oss models. The raw chain of thought is meant for developers and safety practitioners, not general users. The card lists no benchmark numbers.
An open guard that takes a policy as a prompt. Later cards list it as a comparison.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Apache-2.0 weights for both sizes; the 20b fits on 16 GB GPUs per the card.
- Useful 4: Policy-as-prompt classification without retraining; reasoning output for review.
- Notable 4: Later guard cards list it as a comparison (Shieldstral does).
- Novel 3.5: Built on gpt-oss; the policy is supplied at inference time.
Comments