GPT-6 Sol and GPT-6 Luna reach free and paid ChatGPT: cybersecurity and bio-chemical capability rated High

GPT-6 Sol and GPT-6 Luna are now available to ChatGPT users on free and paid plans globally, replacing GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT. The same-named models accessed in Codex and through ChatGPT Work are still previously released versions, so two model sets share one name. The system card separates them by release month: October for the versions released today, September for the versions that remain in use in Codex and Work.

Safety and alignment evaluations are assessed in aggregate, weighing improvements alongside the nature and severity of regressions rather than reporting single gains. Compared with GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT, GPT-6 showed stronger resistance to jailbreaks, including attacks that adapt across multiple turns. In the same comparison, dishonesty, deception and circumvention of guardrails were reduced; where evaluations showed regressions, manual review of failures and adversarial red-team testing found the disallowed responses were generally low severity.

Under the Preparedness Framework, this October release of GPT-6 Sol and GPT-6 Luna is treated as High capability in both Cybersecurity and Biological and Chemical domains. Neither model reaches the High threshold in AI Self-Improvement. Based on that assessment, both October models carry the same set of safeguards detailed in the GPT-5.6 System Card for GPT-5.6 Sol and GPT-5.6 Luna.

Safety and capability evaluations use different reasoning settings, one for everyday use and one for an upper bound. Safety evaluations such as disallowed content and mental health measure performance at the lowest reasoning deployment settings, to capture the vast majority of usage. Capabilities assessments evaluate the models at maximum reasoning effort to get an upper bound of capabilities.

The release covers GPT-6 Sol and GPT-6 Luna in ChatGPT on free and paid plans globally; Codex and ChatGPT Work still run the September versions and are outside this replacement. Safety evaluations target behavior at the lowest reasoning deployment settings, while capability evaluations target an upper bound at maximum reasoning effort, so each conclusion applies only to its own setting.

The system card describes safety and alignment evaluations as an aggregate weighing of improvements and regressions; where regressions appeared, manual review and adversarial red-team testing judged the disallowed responses generally low severity, without a quantified magnitude. Both models are rated High in cybersecurity and bio-chemical domains and neither reaches the High threshold in AI Self-Improvement.