OpenAI updates GPT-6 Luna with shorter responses and stronger safety scores

OpenAI released an updated GPT-6 Luna on October 29 with more concise responses and less overly agreeable language, as detailed on its Deployment Safety Hub. The model uses the same training methods, data types, and safeguards as the October 7 version.

OpenAI continues to classify GPT-6 Luna as High for cybersecurity and biological and chemical capabilities under its Preparedness Framework. The updated model recorded better safety scores for users under 18 in sexual content, eating disorders, emotional reliance, and gore evaluations.

The company also reported higher observed resistance to multiturn jailbreak attempts at every tested attacker budget, although infrastructure problems ended the test early. GPT-6 Luna reached 99.91% robustness in instruction hierarchy evaluations.

Health benchmark results declined from the October 7 model. Its HealthBench score fell from 50.4 to 48.2, while HealthBenchHard dropped from 27.2 to 22.8.