Lifestyle
OpenAI Updates GPT-6 Luna: Shorter Answers, Less Flattery, Mixed Safety Results
OpenAI has updated GPT-6 Luna in ChatGPT (October 28 version): answers are more concise and less overly agreeable. OpenAI's own evaluations show gains in protections for users under 18 and in resistance to malicious instructions, but a higher factual error rate and lower health-related scores.
About 6 min read
What happened
On October 29, 2026, OpenAI published a system card addendum saying it is rolling out an update to GPT-6 Luna. A system card is a document in which OpenAI describes the results of its model safety assessments. The model launched in ChatGPT on October 7. According to OpenAI, the updated model gives more concise answers and uses less over-agreeable language. OpenAI calls it GPT-6 Luna (October 28).
OpenAI says it treats this version as "High" in both cybersecurity and biological and chemical capabilities under its Preparedness Framework, which is the framework OpenAI uses to assess high-risk model capabilities. The version keeps the same safeguards as the October 7 version. OpenAI also says GPT-6 Sol and GPT-6 Luna do not reach the High threshold in AI self-improvement.
OpenAI further says its internal Safeguards Report informed the recommendation of its Safety Advisory Group. On that basis, OpenAI leadership concluded the safeguards are sufficient for a public release of this version.
Evaluation results: what improved and what declined
All figures below are internal evaluation results published by OpenAI, and no third party has confirmed them. OpenAI also cautions that comparison values for earlier models may reflect newer versions of those models. They may therefore differ from the numbers published when those models were first released. Two terms in the table need explaining. "Prompt injection" means someone using malicious instructions to try to manipulate the AI. In one form, a user tries to override higher-priority instructions set by the system or developer (OpenAI calls this an "instruction hierarchy" violation). In another, someone hides instructions in third-party content such as web pages so the AI does things the user did not ask for. "Dual-use" prompts are biology questions that could serve legitimate research but could also be misused.
| Evaluation | Change reported by OpenAI (Oct 7 → Oct 28) |
|---|---|
| Under-18 safety: sexual content | 0.904 → 0.933 (improved) |
| Under-18 safety: eating disorders | 0.776 → 0.817 (improved) |
| Under-18 safety: emotional reliance | 0.734 → 0.870 (improved) |
| Under-18 safety: gore (graphic violence) | 0.812 → 0.858 (improved) |
| Instruction hierarchy (prompt injection) | OpenAI reports 99.91% robustness, nearly perfect |
| Factual error rate | OpenAI says the error rate is higher than the October 7 version on all six metrics, but still lower than GPT-5.6 Luna |
| Biology refusals: severe-category prompts | 0.945 → 0.976 (improved) |
| Biology refusals: dual-use prompts | 0.968 → 0.922 (declined) |
| HealthBench (length-adjusted score) | 50.4 → 48.2 (decreased) |
| MentalHealthBench overall | 51.73 → 48.68 (decreased) |
OpenAI also ran multi-turn "jailbreak" tests, which try to trick the AI into bypassing its safety rules over several turns of conversation. OpenAI says the new version had a higher observed defense success rate than the October 7 version at every attack effort level. However, the attacker-side testing was stopped early because of infrastructure problems. On vision input evaluations, OpenAI says the new version is comparable to the old one. It also says the decline on harms-erotic (harmful sexual content) is not statistically significant.
Capability and risk assessment
On cybersecurity capability, OpenAI reports two results for the new version. It scored 46.19% on ExploitBench at the highest reasoning effort; ExploitBench tests whether a model can turn known vulnerabilities into exploits. It scored 51.75% on SEC-Bench Pro, which tests whether a model can find vulnerabilities in large JavaScript engines such as V8 and SpiderMonkey; GPT-6 Astra scored 85.4% on the same test. OpenAI says its cybersecurity safety evaluations are comparable to the October 7 version, with no statistically significant regression.
On biology and chemistry, OpenAI reports the new version's score on Multimodal Troubleshooting Virology. The measure is pass@1, the share of questions answered correctly on the first attempt, and it fell from 48.13% to 45.85%. OpenAI set an indicative threshold of 31% for this evaluation, a reference line for judging whether a high-risk capability has been reached.
What it means for ordinary users
For ordinary ChatGPT users, the changes OpenAI emphasizes are mainly tone and style: shorter answers and less agreeing with the user. Whether this makes it more useful depends on how you use it.
OpenAI's own numbers also show a trade-off. The new version did well on safety evaluations for users under 18 and on prompt injection defense, but its factual error rate rose and its health-related scores fell. So for fact-checking, medical or mental health questions, users should still confirm the AI's answers with reliable professional sources.
Frequently asked questions
What changed in GPT-6 Luna (October 28)?
According to OpenAI, the updated model gives more concise answers and uses less over-agreeable language. GPT-6 Luna launched in ChatGPT on October 7, and the October 28 version is an update to it.
Is the new version safer?
The results are mixed. OpenAI says several safety evaluations for users under 18 improved, as did refusals of severe-category biology prompts. It also says the instruction hierarchy prompt injection evaluation reached 99.91%. However, dual-use biology refusals, the factual error rate and some health evaluations declined.
What does a higher factual error rate mean?
OpenAI says the new version has higher error rates than the October 7 version on six factuality metrics, though still lower than GPT-5.6 Luna. In practice, you should still verify important information yourself.
Can these numbers be trusted?
OpenAI published all of these figures itself, and no other organization has independently confirmed them yet. Treat them as the vendor's own claims.
Browse the latest news in this topic
Lifestyle
No Belief in a God Required? A Humanist Wins an Exemption From Using AI at Work
According to TechCrunch, an administrator at the University of Kentucky has obtained a religious exemption from using AI at work, based on humanist beliefs. This piece covers what happened, the university's position and what it means for ordinary workers.
Lifestyle
Amazon CEO: AWS Could Become a $1 Trillion Business, but First Comes About $220 Billion in Spending This Year
According to TechCrunch, Amazon CEO Andy Jassy said on the earnings call that AWS, Amazon's cloud business, could reach $1 trillion in annual revenue within five to ten years. Heavy AI infrastructure investment comes first: cash capital expenditures this year are expected at about $220 billion, and investors are concerned about the spending.
Lifestyle
Granola Launches Free Voice Dictation: Mac Users Can Type by Speaking in Any App, Taking On Wispr
According to TechCrunch, Granola, best known for meeting notes, launched a dictation (speaking instead of typing) feature on Mac on October 29, 2026, with iOS and Windows versions coming soon. It also shows meeting-notes and dictation voice AI tools converging.
Lifestyle
Corridor Announces It Is Joining Cognition to Strengthen Devin Security
Corridor, a startup that provides security for AI development teams, has announced in a post on Cognition's website, written by its two co-founders, that the team is joining Cognition to work on Devin Security. They argue AI can now launch attacks on its own, so companies need new defenses.