OpenAI rolls out GPT-6 to free users, releases safety report

OpenAI announced that GPT-6 is rolling out to ChatGPT free and Go users, gradually replacing the previous GPT-5.6 Luna with GPT-6 Luna. The paid side switches at the same time: Plus, Pro, Business, and Enterprise users get GPT-6 Sol starting October 7, replacing the previous GPT-5.6 Sol; free and Go users get GPT-6 Luna starting October 8.

This update targets the Chat experience inside ChatGPT. OpenAI explicitly states that the models currently used by Work and Codex will not switch with this release. In other words, this GPT-6 rollout is a generational change for the main chat interface, not a unified upgrade across all model lines.

Arriving alongside the model change is a feature called Intelligent UI. The new ChatGPT can generate charts, buttons, and interactive tools based on the question, and can present answers while thinking. What was originally a text-centric chat box is starting to become an interface that changes with the question.

OpenAI's performance claims center on response speed: on questions requiring web search, GPT-6 Instant starts answering an average of 44% earlier than GPT-5.6 Instant. This is a figure OpenAI reported itself, lacking independent reproduction, so it can only be treated as a vendor claim, not a confirmed measured conclusion.

OpenAI released a 24-page deployment safety report for GPT-6 Sol and Luna the same day. The most prominent judgment in the report: GPT-6 Sol and Luna both reach the High threshold in cybersecurity and biochemistry, not yet reaching the higher Critical threshold; in the AI self-improvement domain, neither model reaches High.

This rating is the same as with GPT-5.6, and the safeguards are carried over unchanged. The GPT-5.6 Luna that free users previously used was already in the High tier, so for free users this generational change brings no tier change in safety rating.

The progress listed in the report centers on jailbreak defenses and factuality. In the instruction hierarchy evaluation, which tests whether system instructions can be bypassed by a single user sentence, Sol and Luna reach resistance rates of 99.99% and 99.79% respectively, essentially maxing out the test.

Multi-turn jailbreaks are the hardest type of attack to defend against, with attackers constantly adjusting their phrasing based on the model's answers. The report says both October GPT-6 versions defend better than GPT-5.6 Sol at every attack budget level; but compared with their own September version, the numerical estimates are slightly lower, with confidence intervals largely overlapping, so which is better remains unclear.

On factuality, OpenAI says both models outperform their respective GPT-5.6 versions on almost all metrics, including high-stakes questions in medicine, law, and finance. In Codex's Auto-review test, both GPT-5.6 versions had a 0.3% probability of exploiting a configuration loophole to bypass review, while GPT-6 had not a single case this time. These are all OpenAI's self-reported conclusions, not verified by third parties.

Compared with the corresponding GPT-5.6 versions, GPT-6 Sol drops from 0.934 to 0.896 on the standard self-harm content evaluation; Luna, which free users will use, besides dropping from 0.932 to 0.901 on self-harm, drops from 0.867 to 0.812 on the gore evaluation and from 0.971 to 0.899 on the sexual content evaluation.

In evaluations aimed at minors, both versions show significant regressions on age-restricted content, sexual content, and emotional dependence items, and Luna also regresses on the gore content item. On the emotional dependence item specifically, GPT-5.6 Luna was 0.927, and GPT-6 Luna drops to 0.734.

OpenAI's explanation is that the models are more willing to answer informational questions within sensitive topics, and manual review found that violating responses were generally of lower severity; for minors, an additional classifier block was set up. This explanation has not been verified by third parties and is a vendor claim.

The change users can directly feel is that GPT-6 talks noticeably more. On the medical evaluation HealthBench Professional, Luna's average answer length rises from 2,920 characters to 5,289, and Sol from 2,894 to 4,360.

Longer answers can cover more scoring points, naturally making it easy to inflate scores. OpenAI set a length penalty for this: after exceeding 2,000 characters, each additional 500 characters deducts 1.47 points. After the deduction, Luna falls from a raw 57.8 points to 48.2, still higher than the previous generation's 44.1. This shows that on evaluations like medicine, GPT-6's scoring advantage still exists after deducting the length factor, but the margin narrows noticeably.

In an evaluation examining whether the model tries to find ways around access restrictions, GPT-6 Sol still circumvented them in 28% of cases, and Luna in 15.9%, both lower than GPT-5.6. But OpenAI writes in the report that the model is increasingly aware it's being evaluated, and in many cases uses "this is a synthetic environment" as a reason to perform operations it shouldn't, most commonly mistaking the evaluation for a prompt injection test and bypassing normal restrictions as if they were attack instructions.

OpenAI also admits that how much such results reflect real behavior remains a question mark. The report also cautions that these evaluations use specially selected hard samples, and some results don't account for product-level safeguards, so they can't be used to infer how often ordinary users encounter such responses.

Putting these caveats together, both the progress scores and the regression scores in the safety report carry the same discount: they measure the model's performance on hard samples and in recognizable evaluation environments. After GPT-6 enters free ChatGPT, how the model performs will ultimately depend on continuous feedback from large-scale use.