gpt-6 luna update: Less Flattery, Lower HealthBench

The gpt-6 luna update is OpenAI's attempt to make its ChatGPT model talk less and flatter less. In a system card addendum published October 29, OpenAI said the updated GPT-6 Luna, first released in ChatGPT on October 7, gives more concise responses and cuts overly agreeable language. The fine print is the interesting part: some safety scores went up, while health scores and factual accuracy slipped.

What actually happened

OpenAI published the addendum on its Deployment Safety Hub. The headline change is style: shorter answers and less sycophancy. OpenAI says it uses the same types of data and training as the October 7 card, and the model stays under the same Preparedness Framework classification.

That means "High" for cybersecurity and for biological and chemical capability, but not High for AI self-improvement. OpenAI says the existing safeguards are sufficient for this release. Nothing about the gpt-6 luna update changes which risk tier the model sits in.

The scorecard, good and bad

On the plus side, OpenAI's under-18 evaluations moved up across several categories. Emotional reliance jumped from 0.734 to 0.870, eating disorders from 0.776 to 0.817, and gore from 0.812 to 0.858. OpenAI reports 99.91% robustness on its instruction-hierarchy prompt injection test, and a Pareto improvement on its challenging-prompts safety chart.

On the minus side, health took a hit. HealthBench dropped from 50.4 to 48.2 and MentalHealthBench overall from 51.73 to 48.68. Mokaair's breakdown of the card also points out that OpenAI reports a higher factual error rate than the October 7 version on all six of its hallucination metrics, though still lower than GPT-5.6 Luna. Dual-use biology refusals fell from 0.968 to 0.922, while refusals on severe prompts improved.

Shorter and less flattering is the goal. Slightly less careful on health is the trade-off OpenAI disclosed.

On cyber capability, OpenAI lists 46.19% on ExploitBench at the highest reasoning effort and 51.75% on SEC-Bench Pro. For comparison, the card shows GPT-6 Astra at 85.4% on SEC-Bench Pro. All of these are OpenAI's internal numbers; nobody has independently reproduced them yet.

On biology, the card shows Multimodal Troubleshooting Virology at 45.85% pass@1, down from 48.13% for the October 7 version, against an indicative threshold of 31%. TroubleshootingBench slipped from 41.46% to 37.90%. Lower is not automatically safer or worse here; these tests track dangerous know-how, and both versions stay above several of OpenAI's reference lines, which is why the High tier stays.

Mokaair also flags a caveat on jailbreaks: OpenAI says the new version defended better at every attack effort level in multi-turn tests, but attacker-side testing was cut short by infrastructure problems.

Why it matters

Sycophancy has been a running complaint about chat models: they agree too easily and pad answers. A gpt-6 luna update that trims both is aimed straight at that. For heavy users, shorter replies also mean faster reads and fewer tokens.

Luna is not a niche model either. OpenRouter's public rankings, with usage data through October 29, list GPT-6 Luna fourth by tokens processed at about 11.4 trillion, up about 86%. Behavior changes at that scale ripple into apps and agents built on top of it.

If you have followed the family, our gpt-6 sol and gpt-6.1 terra pieces cover the bigger siblings. Luna is the everyday ChatGPT workhorse, so tone tweaks here touch more people than any benchmark headline.

What to do as a reader

If you use ChatGPT for quick answers, expect tighter replies and less "great question!" energy. If you relied on long explanations, ask for detail explicitly.

Be extra careful with health, medical, or mental-health questions. OpenAI's own numbers show those scores dipped in this version. Cross-check anything important with qualified sources, as you should with any chatbot.

If you build on the API or run agents, retest prompts that depend on verbosity or agreeable tone. Concise defaults can break parsers, summaries, or flows tuned to the old style.

Sources: OpenAI system card addendum, Mokaair breakdown, OpenRouter rankings.