What Intelligent UI actually does
OpenAI says it trained GPT-6 to compose responses from text, visuals and interactive elements, choosing how they fit together based on the question. Responses can include graphics, tappable buttons, forms, charts and interactive experiences you can use directly in the conversation.
The format follows the question. A comparison might work best side by side, and an explanation might work better as an interactive diagram. When plain text is the most useful answer, ChatGPT can still give you that.
OpenAI's examples fall into three groups:
- Everyday answers. A cooking timeline can sit beside a recipe, and road-trip stops can be plotted on a map.
- Learning. The announcement says users can explore how something works, change an input to see what happens, or work through a problem step by step.
- Tools on request. OpenAI's examples are a savings calculator, a bill splitter for a dinner with friends, and a game you can play in the conversation.
OpenAI's help documentation says ordinary text answers are still possible, and no special prompt is required. So you don't have to learn any magic words. The model decides when a UI is worth building.
How it works under the hood
OpenAI describes two pieces. It built a library of native, streamable components and a compiler that processes the interface as the model generates it. The library gives responses a familiar look and leaves the model room to arrange the pieces. The compiler lets the interface appear progressively, so the UI doesn't wait for the full response to finish.
This is a constrained approach. The model assembles a vetted set of building blocks. It doesn't write arbitrary front-end code on the fly. In my reading that's a sensible trade-off. It keeps things consistent across web and mobile, and it limits the odd rendering problems that free-form generated interfaces can have. That's analysis, not something OpenAI states.
OpenAI is also candid about the limits. It says there's still work ahead to improve the model's design judgment and expand what it can create. Expect some rough edges.
Answering while still thinking
The second change is less visible but may matter more in daily use. OpenAI says ChatGPT can now interleave thinking with answering. The model builds an answer across several partial responses, each adding useful information, and OpenAI says the finished answer should be as cohesive and factual as one written all at once.
OpenAI's own numbers:
- For questions that need web search, GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant, per OpenAI's announcement.
- In an internal evaluation of everyday agentic tasks, GPT-6 Extra High begins answering in the same time as GPT-5.6 Medium. OpenAI says it scores better overall than GPT-5.6 Extra High.
- In an internal evaluation of difficult problems, GPT-6 more often addressed the key aspect of the user's question than GPT-5.6 did. OpenAI gave no figure for this.
All of these are OpenAI's internal results. None has been independently verified, and they don't guarantee that your particular queries will be faster or better. Time-to-first-token improvements are real, but "starts sooner" isn't the same as "finishes sooner."
Rollout and plans
The timing is as follows:
- October 7: the rollout began for Plus, Pro, Business and Enterprise users in the Chat tab.
- October 8: the rollout expands to Free and Go tiers.
- Models: paid tiers get GPT-6 Sol, and Free and Go get GPT-6 Luna.
- Enterprise: access depends on workplace admin settings.
- Scope: the release covers the Chat experience only. The models behind Work and Codex are unchanged.
OpenAI's system card makes the same point from the safety side. The October versions are the ones in Chat. Users of GPT-6 Sol and Luna in Codex and ChatGPT Work are still on the earlier September versions.
Windows users: check these caveats
OpenAI's Help Center covers the practical limits, and some of them affect Windows and enterprise users directly.
- The older desktop apps don't support it. The older ChatGPT desktop apps for macOS and Windows do not support these capabilities; use ChatGPT on the web instead. If your team lives in the Windows desktop client, the new interface won't appear there. Use a browser tab.
- Mobile apps need updating. OpenAI says to update to the latest available version.
- It isn't available at Pro effort. The capabilities apply to GPT-6 at Instant through Extra High, and aren't available with Pro effort, which remains powered by GPT-6 Astra.
- No separate quota. Intelligent UI has no separate usage quota, and existing model, tool and plan allowances still apply.
- Model menu changes. The help article says GPT-6 replaces the "Latest" entry in the model menu. Existing Latest selections move to GPT-6 at the same effort, and explicit GPT-5.6 Sol selections stay on GPT-5.6.
Turning down the visuals
If you'd rather have fewer widgets, on ChatGPT on the web:
- Open Settings.
- Turn off Layout and visuals.
Don't expect a full classic mode. Some visual elements may still appear, and the setting does not switch the entire answer experience to a classic or text-only mode.
What IT admins should watch
For managed environments, the useful points are:
- Enterprise availability is tied to workspace admin settings. The Help Center says Enterprise customers may keep seeing the existing model picker until the new controls reach their workspace.
- Business and Enterprise admins can control which models workspace members can use. If users report missing GPT-6 or missing interactive answers, check model-access permissions first.
- The Windows desktop client gap is a support-desk item. Users who expect the new interface in the old app should be pointed to the web version.
- Interactive tools generated in a chat are new surface area. OpenAI hasn't described data-handling differences for them in the material I reviewed, so existing AI-use policies are the place to start.
Safety claims
OpenAI says GPT-6 builds on safety advances from its Astra model and that it updated safety training to strengthen protections against high-risk misuse involving cyberattacks, biological threats and violence. It also reports stronger resistance to jailbreaks, particularly attacks that adapt across multiple turns, compared with GPT-5.6 Sol and Luna in ChatGPT.
The system card adds nuance:
- OpenAI treats the October GPT-6 Sol and Luna as High capability in both cybersecurity and biological and chemical domains, and below the Critical threshold in cybersecurity.
- It reports statistically significant regressions on some safety categories. Sol regressed on standard self-harm. Luna regressed on standard self-harm, gore and sexual content.
- OpenAI says its manual review found these violations borderline but generally safe, and that it identified no high-severity risks in adversarial red teaming.
These are OpenAI's own assessments, and the company says they are weighed in aggregate against regressions. For readers in security, the takeaway is that the improvements are real by OpenAI's account but are not a clean sweep.
The bigger picture
OpenAI's closing argument is that software should adapt to people rather than people adapting to software. That is an ambitious pitch. It also echoes a long-running idea in computing: dynamic, task-shaped interfaces instead of fixed menus.
Still, it's early. Interface quality depends on the model's design judgment, which OpenAI admits is unfinished. Generated tools such as calculators are only as trustworthy as the logic the model writes for them. If a bill splitter gets the maths wrong, a prettier interface won't help. Check outputs that matter, as you would with any AI answer.
Bottom line
- GPT-6 with Intelligent UI is rolling out in ChatGPT's Chat tab. Paid plans got it October 7 and Free and Go follow October 8.
- Answers can include interactive visuals, forms, charts and small tools, and plain text remains an option.
- Windows users should use ChatGPT on the web, because the older desktop app doesn't support the feature.
- Speed and quality claims, including the 44% faster start on web-search questions, are OpenAI's internal results.
- Enterprise access depends on admin settings, and Work and Codex are not part of this release.