gpt-oss-safeguard

2025-10-29 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI releases gpt-oss-safeguard-120b and 20b in preview, open-weights reasoning models that classify content according to a safety policy provided by the developer at inference time.

What changes

  • Fine-tuned versions of gpt-oss, under the Apache 2.0 license, for safety classification tasks.
  • The model receives two inputs, a policy and the content to classify, and returns its conclusion along with its reasoning.
  • The policy is not trained into the model: it can be changed without retraining a classifier.
  • Derived from the internal Safety Reasoner tool, developed with ROOST, which is launching a community around open safety models.

### Specifications - Two sizes: gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, weights downloadable on Hugging Face. - Chain of thought that the developer can inspect to explain each decision. - Context, maximum output and knowledge cutoff: not detailed in the post (inherited from gpt-oss). ### Pricing Open weights under the Apache 2.0 license, with no OpenAI pricing. ### Benchmarks Scores published by the vendor, in qualitative form: on an internal multi-policy evaluation, gpt-oss-safeguard and the internal Safety Reasoner outperform gpt-5-thinking and gpt-oss in accuracy. On the 2022 moderation dataset, gpt-oss-safeguard is slightly above the other models tested, with no statistically significant gap with the Safety Reasoner. The numerical values are in the technical report. ### Key takeaways Useful for platforms that need to apply their own moderation rules, with a compute cost and latency higher than those of a conventional classifier.

Links

Tags : Open Source, LLM, OpenAI, Sécurité, gpt-oss, Modération