OpenAI releases gpt-oss-safeguard

OpenAI has released a research preview of gpt-oss-safeguard, an open-weight reasoning model for safety classification tasks. The model is available in two sizes, gpt-oss-safeguard-120b and 20b, and can be freely used and modified under the Apache 2.0 license.

This model directly interprets developer-provided policies at inference time through reasoning. Using a Chain-of-Thought (CoT) approach, it shows the process by which the model arrives at its conclusions, allowing developers to review the model's rationale.

While traditional safety classifiers (e.g., the Moderation API) learn from thousands of examples to indirectly infer policies, gpt-oss-safeguard takes the policy itself as input. This makes it particularly useful in the following situations:

• When new types of harm emerge and policies need to be changed quickly
• When a domain is so nuanced that small classifiers struggle to handle it
• When there aren't enough samples to train a high-quality classifier for a specific risk
• When explainable, high-quality labeling matters more than latency