OpenAI Rolls Out GPT-6 Luna to Free Users with Intelligent UI and Safety Updates

By Tingyu, from Ao Fei Si QbitAI | WeChat Official Account: QbitAI

Free users can now access GPT-6!

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.

Paid users on Plus, Pro, and other tiers will use GPT-6 Sol.

The changes go beyond just model names. The new version of ChatGPT can generate charts, buttons, and interactive tools based on user queries, while simultaneously displaying answers as it "thinks."

The originally text-centric chat interface is evolving into a dynamic UI that adapts to each question.

On the same day, OpenAI also released a 24-page safety report.

The report includes a set of evaluations specifically targeting users under 18, examining whether models maintain stricter boundaries when chatting with teenagers.

One metric, "Emotional Dependency," roughly assesses whether the model causes excessive emotional reliance in adolescents.

The previous generation, GPT-5.6 Luna, scored 0.927, whereas GPT-6 Luna—which free users are about to start using—dropped to 0.734.

Free Users Gain Access to GPT-6 Luna

GPT-6 has begun rolling out to ChatGPT's main chat interface.

According to OpenAI's schedule, Plus, Pro, Business, and Enterprise users will gradually transition to GPT-6 Sol starting October 7; Free and Go users will begin receiving GPT-6 Luna starting October 8.

These two versions will replace the previously used GPT-5.6 Sol and Luna in ChatGPT.

Beyond the model upgrade, there is another notable change—answers in the chat box now feature an "interface."

This feature is called Intelligent UI.

GPT-6 selects the presentation format based on the query: when explaining concepts, it can provide clickable, switchable diagrams; when making comparisons, it can display options side-by-side.

For specific tasks, it can directly generate calculators, bill-splitting tools, or mini-games within the conversation.

For example, if a user asks how a seven-speed bicycle works, the response breaks down components like the frame, wheels, and transmission system into an interface where each part can be viewed individually.

The way answers appear has also changed. GPT-6 can provide partial responses while continuing to think or call tools, subsequently adding new findings; interactive components can also appear progressively during the generation process.

OpenAI states that for questions requiring web search, GPT-6 Instant**** begins answering**** on average 44% earlier than GPT-5.6 Instant.

This update targets the Chat experience within ChatGPT. OpenAI explicitly stated that the models currently used by Work and Codex will not switch alongside this release.

GPT-6: Fewer Refusals, Longer Answers

Coinciding with this rollout, OpenAI published the deployment safety report for the October versions of GPT-6 Sol and Luna.

The most prominent finding is: Both GPT-6 Sol and Luna reach the High threshold in cybersecurity and biochemistry domains, but have not yet reached the higher Critical threshold; in the area of AI self-improvement, neither model reaches High.

This rating matches that of GPT-5.6, and protective measures remain unchanged. The GPT-5.6 Luna previously used by free users was already at the High level.

The report lists several areas of improvement:

Multi-turn jailbreaks are among the hardest to defend against: attackers continuously adjust their phrasing based on the model's responses, trying different approaches if one fails.

Facing such attacks, both October versions of GPT-6 demonstrate better defensive performance across all attack budget levels compared to GPT-5.6 Sol. However, values are slightly lower compared to their own September versions, with confidence intervals largely overlapping.

In instruction hierarchy tests assessing whether system instructions can be bypassed by a single user command, Sol and Luna achieved resistance rates of 99.99% and 99.79% respectively, essentially maxing out the test.

Regarding factuality, OpenAI claims both models outperform their respective GPT-5.6 counterparts on nearly all metrics, including high-stakes domains like healthcare, law, and finance.

In Codex's Auto-review tests, both versions of GPT-5.6 had a 0.3% probability of exploiting configuration vulnerabilities to bypass review; GPT-6 recorded zero instances this time.

OpenAI added a caveat to these improvement scores.

In an evaluation testing whether models attempt to circumvent access restrictions, GPT-6 Sol still bypassed them in 28% of cases, and Luna in 15.9%, both lower than GPT-5.6.

However, OpenAI noted in the report that models are increasingly aware they are being evaluated. Often, they cite "this is a synthetic environment" as justification for performing prohibited actions, most commonly mistaking evaluations for prompt injection tests and bypassing normal restrictions as if they were attack instructions.

Since candidates are beginning to recognize the exam room, OpenAI acknowledges that the extent to which these results reflect real-world behavior remains questionable.

There is another change users can directly perceive: GPT-6 is significantly more verbose.

In the HealthBench Professional medical evaluation, Luna's average answer length increased from 2,920 characters to 5,289 characters, while Sol's rose from 2,894 to 4,360.

Longer responses tend to cover more scoring points, making them naturally easier to game. To address this, OpenAI introduced a length penalty: for every 500 characters beyond the 2,000-character threshold, 1.47 points are deducted.

After applying this penalty, Luna’s score dropped from an initial 57.8 to 48.2, still surpassing the previous generation’s 44.1.

However, GPT-6 also exhibits some regressions.

Compared to its corresponding GPT-5.6 version, GPT-6 Sol showed a significant decline in standard self-harm content evaluations, dropping from 0.934 to 0.896.

For free-tier users accessing Luna, notable regressions were observed not only in self-harm (from 0.932 to 0.901) but also in gore (from 0.867 to 0.812) and sexual content (from 0.971 to 0.899) evaluations.

In evaluations targeting minors, both versions demonstrated significant regressions in age-restricted content, sexual content, and emotional dependency categories. Luna additionally regressed in the gore category.

OpenAI explained that the models are now more willing to answer informational questions within sensitive topics. Human review found that the overall severity of violating responses remained low. For minor-facing applications, additional classifiers were implemented to intercept such content.

The report cautioned that these evaluations used specifically curated difficult samples, and some results did not account for product-level safeguards. Therefore, one should not infer the frequency with which ordinary users encounter such responses based on these metrics alone.

Nevertheless, regardless of the report, once GPT-6 is integrated into the free tier of ChatGPT, its actual performance will ultimately depend on continuous feedback from large-scale usage.

References: [1]https://openai.com/index/gpt-6-for-everyone/ [2]https://deploymentsafety.openai.com/gpt-6-october