What happened
OpenAI published an August 2026 system-card update for GPT-5.6 Sol and GPT-5.6 Luna as they roll out to ChatGPT's paid and free tiers respectively, distinguishing these 'August' checkpoints from the July versions still in use in Codex and ChatGPT Work. Under its Preparedness Framework, OpenAI classifies this August release as 'High capability in both Cybersecurity and Biological and Chemical domains' — consistent with the original July GPT-5.6 System Card's classification — while confirming neither model reaches the High threshold for AI Self-Improvement. The substantive new disclosure is the introduction, for the first time, of dedicated U18 (under-18) evaluations measuring model behavior against teen-specific safety standards, alongside updated Production Benchmark results for disallowed content, mental-health-sensitive prompts, and capability assessments run at maximum reasoning effort. OpenAI notes GPT-5.6 Sol performs 'comparably' to the prior production model (GPT-5.5 Instant) on most disallowed-content categories, with exceptions in gore and disallowed sexual content.
Why it matters
This is the first frontier-lab system card to publish dedicated under-18 safety evaluations, setting a disclosure precedent that regulators and child-safety advocates will likely point to when assessing other labs; CISOs and trust & safety leads should note the reaffirmed High-capability classification in cybersecurity and bio/chem domains as it affects deployment risk posture for any enterprise integrating GPT-5.6.
Action needed
Compliance and Trust & Safety teams should review the U18 evaluation methodology and disallowed-content benchmark deltas against their own deployment context, particularly for consumer-facing products serving minors.