OpenAI Publishes GPT-6 Safety Assessment, Rates Models 'High Capability' Risk

OpenAI has published safety assessment material for its GPT-6 models on its deployment safety site. The excerpt covers how the models were trained, how they were rated for biological and chemical capability, and how they performed on safety evaluations in biology and cybersecurity. It names two models, GPT-6 Sol and GPT-6 Luna, and refers readers to a separate GPT-6 Astra system card for full descriptions of the evaluations.

On training, OpenAI says the GPT-6 models learned from varied datasets that were run through a data processing pipeline, partly to cut down personal information. As with its other reasoning models, they are trained through reinforcement learning to think before answering. They can generate a long internal chain of thought, test different strategies and notice their own errors. OpenAI says this reasoning helps the models follow its guidelines and policies, give more helpful answers and better withstand attempts to get around safety rules.

On capability ratings, OpenAI says both GPT-6 Sol and GPT-6 Luna are handled as High capability within the biological and chemical domain. GPT-6 Sol was also tested against Critical capability evaluations, and the results stayed below the indicative Critical thresholds. Separate Critical testing was judged unnecessary for GPT-6 Luna, because its scores on all High capability evaluations trailed GPT-5.6 Sol.

On safety evaluations, the October versions of GPT-6 Sol and Luna show substantially better results on severe and dual-use biology prompts than the August GPT-5.6 versions, measured on model responses alone without production safeguards. In cybersecurity, scores are broadly similar to the earlier models, with refusal described as one layer of a defense-in-depth stack.