On October 8, 2026, OpenAI officially launched its new flagship model GPT-6 to ChatGPT users worldwide, gradually replacing the previous GPT-5.6 Sol and GPT-5.6 Luna.
The highlight of this update is not just a simple model parameter upgrade, but the introduction of a new interactive layer called "Intelligent UI" in ChatGPT for the first time - AI answers are no longer limited to pure text, but can autonomously choose the appropriate presentation form according to the type of question. Images, charts, buttons, forms, and even directly operable small interfaces can be generated in real-time in the dialogue window.
This GPT-6 push is mainly aimed at the Chat experience scenario of ChatGPT. OpenAI has explicitly stated that the models currently used by Work and Codex will not switch with this release, and these two product lines are still running the previously released versions
Push rhythm and dual model strategy
The upgrade covers a wide range. Plus, Pro, Business, and Enterprise users have been receiving updates since October 7th, while free and Go users will follow up after October 8th.
The models corresponding to different packages are differentiated: paying users run GPT-6 Sol, while free and Go users use GPT-6 Luna. Both models have been specifically optimized for daily conversation experience, and not the same set of weights directly applies to all users.
This dual model strategy originates from the GPT-6 series product matrix released by OpenAI on September 22nd. At that time, OpenAI officially launched GPT-6 Sol and GPT-6 Luna, which absorbed multiple capability upgrades from GPT-6 Astra and permanently lowered API prices by 50%.
In terms of specific pricing, the input unit price of GPT-6 Sol per million tokens has been reduced from $4 to $2, and the output unit price has been reduced from $20 to $10; The input unit price of GPT-6 Luna per million tokens has decreased from 0.2 US dollars to 0.1 US dollars, and the output unit price has decreased from 1.2 US dollars to 0.5 US dollars.
Thus, OpenAI has formed a three-layer product matrix consisting of its flagship model GPT-6 Astra, mid-range model GPT-6 Sol, and lightweight model GPT-6 Luna.
It is worth noting that GPT-6 Sol performs well in professional workflow testing. In the highest inference setting of the Automation Bench, GPT-6 Sol achieved a score of 33.2%, higher than Claude Opus 5's 26.9% in maximum effort mode, and the single task cost was only about one ninth of the latter. In the Agents' Last Exam test, GPT-6 Sol achieved a score of 56.4%.
Intelligent UI: From "Reading Answers" to "Operating Answers"
What truly changes the interactive experience is the new interface layer of intelligent UI. In the past, the way ChatGPT answered questions was similar to writing an article - outputting a paragraph of text in one breath. Now, GPT-6 will first determine the nature of the problem and then decide which form to present the answer in.
According to OpenAI's official blog, GPT-6 has been specially trained to use text, visual elements, and interactive components to organize answers and select combinations of these elements based on user questions. The response forms supported by intelligent UI include graphics, clickable buttons, forms, charts, and interactive experiences that can be directly used in conversations. The choice of format depends on the content asked by the user - comparative questions are suitable for displaying side by side, explanatory concepts are easier to understand through interactive diagrams, and simple questions can still be answered directly in text.
In specific application scenarios, when users explore how aircraft wings generate lift, the system can generate multiple schematic diagrams to visually display air flow, pressure differences, and lift formation mechanisms. When dealing with real-life scenarios, customized visual aids such as recipe images, bicycle mechanical schematics, multi day hiking maps, and personal savings calculators can be generated. In more practical scenarios, users can generate calculators, bill sharing tools, and even create actionable mini games directly in the chat window with just one sentence from ChatGPT.
OpenAI product manager Aarush Selvan said that ChatGPT has long been centered around text interaction, but the company wants to help users truly accomplish tasks, whether it's planning recipes, arranging trips, or learning new knowledge. He believes that the most valuable answers are often not just words, but comprehensive experiences that can help users understand and manipulate information.
Technically, the intelligent UI is built on a native component library that can be streamed and a compiler that synchronously processes interface elements during the GPT-6 response generation process. This means that the interactive component will gradually display as the answer is generated, rather than presenting it all at once after the entire answer is generated.
OpenAI also has boundaries: the intelligent UI will not appear in every answer, simple questions will still be answered in plain text, and only switch forms when it is determined that graphics or interactive elements are truly needed. Users can also adjust the display ratio of visual elements in their replies according to their preferences. If they prefer traditional text replies, they can also reduce or turn off visual elements.
Response speed: Speaking while thinking
The pain point of response speed, which has long plagued inference models, has also been optimized in a targeted manner. In the past, inference models needed to complete the entire internal thinking before outputting the first word, which required users to wait for a long time. GPT-6 adopts the mechanism of "speaking while thinking" - while reasoning advances in the background, the answers have already started streaming out, and as thinking deepens, the previous content is supplemented and revised layer by layer, significantly reducing the sense of waiting.
According to data released by OpenAI, the average waiting time for GPT-6 Instant to receive a question and spit out the first answer for questions that require online search has been reduced by 44% compared to GPT-5.6 Instant.
In addition, in the internal evaluation of high-value daily agent tasks, the response time of GPT-6 Extra High is roughly equivalent to GPT-5.6 Medium, but the overall score is higher than GPT-5.6 Extra High.
Model Foundation: Network Search and Security Capability Enhancement
In addition to the interaction layer and speed optimization, GPT-6 has also made multiple improvements in its basic capabilities.
In terms of online search, the new model can more accurately determine when to search and how to use search results to support answers, which has improved the credibility of answers.
In terms of security, GPT-6 integrates GPT-6 Astra's security technology improvements, strengthening protection against abuse in the fields of networking, biology, and violence. According to the deployment security report released by OpenAI, GPT-6 Sol and Luna have both reached the high capability threshold in the fields of network security and biochemistry, and have not yet reached the critical level; In the field of AI self-improvement, neither model has reached a high level.
In multiple rounds of jailbreak attack testing, both versions of GPT-6 performed better in defense than GPT-5.6 Sol in each attack budget. In the instruction level evaluation, the rates of Sol and Luna resisting system instructions being bypassed reached 99.99% and 99.79%, respectively. In terms of facts, OpenAI claims that both models outperform their respective GPT-5.6 versions in almost all metrics, including areas with high cost of incorrect answers such as healthcare, law, and finance.
In a specialized safety evaluation targeting users under 18 years old, the "emotional dependence" indicator of GPT-6 Luna decreased from 0.927 in the previous generation GPT-5.6 Luna to 0.734, indicating that the new model has improved in avoiding excessive emotional dependence among teenagers.
The release of GPT-6 marks a step forward for ChatGPT from a tablet that only knows how to write to an intelligent workbench that can decide how to present itself and speak while thinking. Intelligent UI liberates AI answers from pure text to interactive interfaces, enabling some operations that originally required specialized software or web pages to be completed through natural language dialogue within ChatGPT. During the continuous evolution of video generation models such as Veo and Sora, OpenAI has chosen to make a breakthrough in the interaction layer, reflecting that competition among large models is shifting from "answer quality" to "task completion efficiency".
