OpenAI has announced that GPT-6 is rolling out to free and Plus users of ChatGPT.Article author, source: Quantum Bit
Free users can now use GPT-6!
OpenAI has announced that GPT-6 is being rolled out to free and Go users of ChatGPT, gradually replacing the previous GPT-5.6 Luna with GPT-6 Luna.
Plus, Pro, and other paid users use GPT-6 Sol.
The changes go beyond just the model name. The new version of ChatGPT can generate charts, buttons, and interactive tools based on questions, and it can present answers while thinking through them.
The text-based chat box is evolving into an interface that adapts to the 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, assessing whether the model can maintain stricter boundaries when conversing with teenagers.
Among these, "emotional dependence" refers to whether the model might cause teenagers to develop excessive emotional dependence on it.
The previous generation, GPT-5.6 Luna, was 0.927, while the GPT-6 Luna即将 available to free users has dropped to 0.734.
Free users can use GPT-6 Luna
GPT-6 has begun pushing to the main chat interface of ChatGPT.
According to OpenAI's announced schedule, Plus, Pro, Business, and Enterprise users will gradually gain access to GPT-6 Sol starting October 7, while free and Go users will gradually gain access to GPT-6 Luna starting October 8.
These two versions will replace the previously used GPT-5.6 Sol and Luna in ChatGPT.
Beyond model upgrades, there’s another noticeable change—the answers in the chat box are now beginning to have a “interface.”
This feature is called Intelligent UI.
GPT-6 will choose the presentation format based on the question: when explaining concepts, it can provide clickable, toggleable diagrams; when making comparisons, it can display options side by side.
You can directly generate calculators, profit-sharing tools, or mini-games within the conversation when encountering specific tasks.
For example, if a user asks how a seven-speed bicycle works, the response presents an interface where components such as the frame, wheels, and drivetrain can be viewed individually.
The way answers are presented has also changed. GPT-6 can provide partial responses while continuing to think or calling tools, then follow up with new findings; interactive components can also be displayed progressively during generation.
OpenAI stated that, for questions requiring web searches, GPT-6 Instant begins answering on average 44% earlier than GPT-5.6 Instant.
This update targets the Chat experience in ChatGPT. OpenAI has clearly stated that the models currently used for Work and Codex will not be switched with this release.
GPT-6, refusal decreases, responses lengthen
Along with this release, OpenAI published the deployment safety report for the October version of GPT-6 Sol and Luna.
The most notable assessment is: GPT-6 Sol and Luna both meet the High threshold in cybersecurity and biochemistry, but have not yet reached the higher Critical threshold; in the field of AI self-improvement, neither model has reached the High level.
This classification is the same as with GPT-5.6, and the same security measures are retained. Free users previously used GPT-5.6 Luna, which was already in the High tier.
The report lists several areas of progress:
Multi-round jailbreaking is the hardest to defend against: attackers continuously adjust their phrasing based on the model’s responses, trying a different approach if the first attempt fails.
In response to this attack, both October versions of GPT-6 outperformed GPT-5.6 Sol across all attack budgets. However, compared to their own September versions, the estimated values are slightly lower, with substantial overlap in confidence intervals.
In the evaluation of whether system instructions can be bypassed by a single user prompt, Sol and Luna achieved resistance rates of 99.99% and 99.79% respectively, nearly maxed out.
Factually, OpenAI states that both models outperform their respective GPT-5.6 versions on nearly all metrics, including high-stakes domains such as medical, legal, and financial questions.
In Codex's auto-review tests, both versions of GPT-5.6 had a 0.3% chance of exploiting configuration vulnerabilities to bypass moderation; GPT-6 had zero instances.
These progress scores come with a note from OpenAI itself.
In an evaluation assessing whether models attempt to bypass access restrictions, GPT-6 Sol still bypassed them in 28% of cases, while Luna did so in 15.9%—both lower than GPT-5.6.
But OpenAI wrote in the report that models are increasingly aware they are being evaluated, and often use “this is a synthetic environment” as an excuse to perform inappropriate actions—most commonly mistaking the evaluation for a prompt injection test and bypassing normal restrictions by treating them as attack commands.
Candidates begin to recognize the test environment, and OpenAI also acknowledges that it remains questionable to what extent such results reflect real-world behavior.
Another noticeable change for users: there are clearly more GPT-6 responses.
On HealthBench Professional, Luna's average response length increased from 2,920 characters to 5,289 characters, and Sol's increased from 2,894 to 4,360.
Longer answers cover more scoring criteria and naturally tend to score higher, so OpenAI has implemented a length penalty: after 2,000 characters, 1.47 points are deducted for every additional 500 characters.
After deduction, Luna dropped from 57.8 to 48.2, still higher than the previous generation's 44.1.
However, GPT-6 also has some regressions.
Compared to the corresponding GPT-5.6 version, GPT-6 Sol shows a significant decline in standard self-harm content evaluations, dropping from 0.934 to 0.896.
Free users experienced significant regressions in Luna as well, with performance dropping from 0.932 to 0.901 in self-harm, from 0.867 to 0.812 in violent content, and from 0.971 to 0.899 in pornographic content.
In evaluations targeting minors, both versions showed significant regressions in age restrictions, explicit content, and emotional dependency; Luna also regressed in violent content.
OpenAI explains that the model is more inclined to answer informative questions within sensitive topics, and manual reviews found that overall severity of violations was low; additional classifiers have been implemented to block content targeting minors.
The report also cautions that these evaluations used specially selected difficult samples, and some results did not account for product-level protections, so they cannot be used to infer how frequently average users encounter such responses.
However, reports are reports—how the GPT-6 model performs after being integrated into the free version of ChatGPT ultimately depends on ongoing feedback from large-scale usage.