OpenAI introduces gpt-oss-safeguard for open-weight safety classification

OpenAI released gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, two open-weight models fine-tuned from its earlier gpt-oss release specifically for safety classification and released under the same permissive Apache 2.0 licence. Rather than being trained against a fixed set of categories the way OpenAI’s existing moderation API works, the models take a developer-supplied written policy as input and reason from it at inference time to label content, showing their full chain of thought behind the classification. OpenAI said this let a single model generalise across different products’ moderation needs — content filtering, trust-and-safety labelling, input/output filtering — without retraining for each one, and that developers could adjust reasoning effort between low, medium and high depending on latency and accuracy needs.

The release extended OpenAI’s earlier open-weight gpt-oss models into a specific safety application, and put a downloadable, self-hostable moderation tool into the hands of developers who did not want to send content to OpenAI’s cloud API or rely on categories OpenAI itself defined. It arrived as open-weight releases from OpenAI, Meta and Chinese labs continued to narrow the capability gap with closed frontier models on narrower, task-specific applications even where the largest closed models remained ahead on general benchmarks.