OpenAI Publishes Safety Card for GPT-6 Sol and GPT-6 Luna Rolling Out in ChatGPT

OpenAI published deployment safety documentation for GPT-6 Sol and GPT-6 Luna, the models replacing GPT-5.6 Sol and Luna in ChatGPT across free and paid consumer accounts globally, detailing concrete gains and regressions versus the prior generation.

Details

• Rollout scope: GPT-6 Sol and Luna replace GPT-5.6 versions in ChatGPT free and paid consumer accounts worldwide; GPT-5.6 Sol/Luna (released August 2026) remain in use for Codex and ChatGPT Work
• Security robustness: GPT-6 Sol scored 99.99% and GPT-6 Luna 99.79% on prompt-injection robustness tests; GPT-6 Sol showed higher defender success rates than GPT-5.6 Sol across tested attacker budgets, with stronger resistance to multi-turn adaptive jailbreaks
• Policy compliance: Auto-Review circumvention violations fell to 0% for both GPT-6 models, down from 0.3% for GPT-5.6
• Factuality: significant reduction in response-level errors on high-stakes medical, legal, and financial prompts, plus reduced dishonesty and deceptive behavior toward users
• Regressions: statistically significant increases in responses touching self-harm for both models, greater willingness to engage with gore and sexual content (OpenAI characterizes severity as “generally lower”), and regressions in age-appropriate content handling for users under 18
• Capability classification: both models rated “High” capability in cybersecurity and biological/chemical domains — below the “Critical” threshold — and neither reaches “High” in AI self-improvement
• Caveat: OpenAI notes the evaluations measure baseline model behavior and exclude production-level safety systems such as classifier-based blocking and age-appropriate safeguards that are layered on top in deployment

What happened next

The safety card lands alongside OpenAI’s broader GPT-6 rollout to ChatGPT’s full consumer base, and its mixed results — security and factuality gains offset by self-harm and minor-safety regressions — illustrate the trade-offs OpenAI is now disclosing more systematically as frontier models get faster at resisting attacks but not uniformly safer across every risk category.