OpenAI Flags Jailbreak Gains and Safety Trade-Offs in October GPT-6 Models

OpenAI's October GPT-6 release brings Sol to paid ChatGPT users and Luna to Free and Go users, replacing the earlier GPT-5.6 versions in the consumer product. The accompanying system card says the models resist jailbreaks better overall but show statistically significant regressions on some safety benchmarks, including self-harm for Sol and self-harm, gore, and sexual content for Luna. OpenAI rates both models as High capability in cybersecurity and biological/chemical domains, while describing the reviewed failures as generally low severity and noting additional system-level safeguards.

OpenAI’s October GPT-6 rollout splits by tier: Sol reaches paying ChatGPT customers, while Luna serves Free and Go. These consumer models supersede GPT-5.6 Sol and Luna, though Codex and ChatGPT Work still use September builds. The system card rates both as High capability for cybersecurity and biological/chemical risks, below Critical cybersecurity and below High for AI self-improvement.

Testing found improved jailbreak resistance overall, yet Production Benchmarks showed significant declines for Sol on self-harm and for Luna on self-harm, gore, and sexual content. U18 evaluations also declined on age-restricted material, sexual content, and emotional reliance, with Luna additionally on gore. OpenAI calls reviewed failures generally low severity and cites extra runtime safeguards.