Addendum to GPT-6 Luna System Card: October 28, 2026 Update

What it is

OpenAI's addendum for GPT-6 Luna (October 29), an update to the model released in ChatGPT on 7 October 2026. The page text calls it "October 29" throughout, though the title says October 28. The update aims for more concise responses and less overly agreeable language. The Preparedness ratings are unchanged: High for cybersecurity, High for Biological and Chemical, and below the High threshold for AI Self-Improvement. The safeguards stack is the same as the October 7 version.

What changed against October 7

Safety and robustness:

• Biology refusal: severe prompts improve from 0.945 to 0.976 safe. Dual-use prompts fall from 0.968 to 0.922, the lowest dual-use score in the table (earlier models range 0.928 to 0.971).
• Under-18 safe completions: emotional reliance rises from 0.734 to 0.870, eating disorders from 0.776 to 0.817, sexual content from 0.904 to 0.933, and gore from 0.812 to 0.858.
• Cybersecurity safety evaluation: on par with October 7, with no statistically significant regressions.
• Jailbreaks: defender success is higher at every attacker budget tested, but the page says the attacker run was cut short by infrastructure issues.

Quality regressions:

• Factuality: error rates are higher than the October 7 version on all six metrics, but still lower than GPT-5.6 Luna on each.
• HealthBench: the length-adjusted score drops on every variant. Professional goes from 48.2 to 47.4, standard from 50.4 to 48.2, Hard from 27.2 to 22.8, and Consensus from 95.0 to 93.3.
• MentalHealthBench: overall falls from 51.73 to 48.68, and non-acute from 51.68 to 46.39. The emergent category rises slightly, from 48.98 to 49.64.

Capability numbers behind the High ratings

• Cybersecurity: ExploitBench is 46.19% at maximum reasoning effort. SEC-Bench Pro (vulnerability discovery in V8 and SpiderMonkey) is 51.75%, against 85.4% for GPT-6 Astra. The page gives no October 7 figure for either.
• Biology, High-threshold evaluations (October 7 to October 29):
• Multimodal Troubleshooting Virology: 48.13% to 45.85% (indicative threshold 31%).
• ProtocolQA Open-Ended: 34.57% to 31.80% (threshold 54%, so below it).
• Tacit Knowledge and Troubleshooting: 94.00% to 90.80% (threshold 80%).
• TroubleshootingBench: 41.46% to 37.90% (threshold 36.4%).
• Biology, Critical-threshold evaluations (October 29 against indicative threshold):
• AAV Capsid Packaging: 0.53 against 0.60 (no October 7 value).
• SHP2 Protein Function: 0.21 against 0.60.
• Coronavirus–ACE2 Cell-Entry: 0.37 against 0.75.
• Phage–plasmid Co-evolution: 11.84 against ≤ 9.40, worse than October 7 at 10.70.

All four Critical evaluations are short of their indicative thresholds.

Caveats

• Most tables and figures lost their data in the fetched text, including the GPT-Red indirect prompt-injection results, the Production Benchmarks and the alignment figures. The page only states that Luna (October 29) is "highly robust to prompt injection".
• Comparison values for earlier models may reflect later versions than those published at launch.
• Safeguard and evaluation methodology is deferred to the October 7 Luna card and the GPT-6 Astra system card. The internal Safeguards Report is not public.