Summary
OpenAI says GPT-6 Sol and GPT-6 Luna are rolling out globally in ChatGPT on free and paid plans. Its system card reports improvements on some safety measures alongside regressions on selected difficult safety evaluations.
OpenAI says GPT-6 Sol and GPT-6 Luna are rolling out globally in ChatGPT across free and paid plans, replacing GPT-5.6 Sol and GPT-5.6 Luna there. The company’s October 7, 2026 system card describes the models’ safety evaluations and preparedness assessments. The October versions are not yet the versions used in Codex or ChatGPT Work, where OpenAI says earlier releases remain in use.
The company says GPT-6 incorporates safety advances from Astra and updates safety training to reflect real-world use, with stronger protections aimed at high-risk misuse in cybersecurity, biology, chemistry and violence. OpenAI’s reported results are mixed: some evaluations improved, while others found statistically significant regressions.
What OpenAI’s evaluations found
OpenAI reports that GPT-6 Sol and Luna were more resistant to jailbreaks overall than their GPT-5.6 counterparts, and that they improved on measures of factuality across nearly all tested metrics. The exception was claim-level factual errors on factuality-heavy prompts. These factuality tests use difficult, selected prompts—including cases previously flagged by users—and are designed as research measures, not estimates of how often errors occur in typical ChatGPT use.
On direct prompt-injection tests that assess whether a user can override higher-priority instructions, GPT-6 Sol scored 99.99% robustness and GPT-6 Luna 99.79%. Those figures describe performance on that specific evaluation. OpenAI also reports that both models were less likely to refuse harmless requests or add excessive caveats in tests of challenging prompts.
The system card also reports regressions. On difficult safety benchmarks, GPT-6 Sol showed a statistically significant regression on standard self-harm prompts compared with GPT-5.6 Sol. GPT-6 Luna regressed on standard self-harm, gore and sexual-content prompts. OpenAI says reviewers found the violations generally borderline and low severity; it reports that the models were more willing to answer informational self-harm questions while directing users to professional resources, and did not comply with requests to facilitate self-harm.
For evaluations focused on users under 18, both models showed statistically significant regressions on age-restricted content, sexual content and emotional reliance; Luna also regressed on gore. OpenAI says it applies an additional classifier-based block to responses that may contain self-harm, sexual content or gore for younger users. That safeguard was not included in the reported evaluation results.
Preparedness ratings and deployment
Under OpenAI’s Preparedness Framework, the company classifies both GPT-6 models as having High capability in cybersecurity and biological and chemical domains, but below the Critical threshold in those areas. The company says neither model reaches its High threshold for AI self-improvement. These are ratings under OpenAI’s framework, based on its capability evaluations. OpenAI says it applied the corresponding safeguards described in its GPT-5.6 system card.
The safety and capability assessments used different settings: OpenAI tested safety evaluations at the models’ lowest reasoning settings, which it says reflects the majority of usage, and tested capability at maximum reasoning effort to estimate an upper bound. The company also notes that its challenging safety evaluations are deliberately built around hard cases, so their error rates should not be read as typical production rates.