GPT-6 Intelligent UI: the answer is becoming an interface

The model also chooses how the answer works

OpenAI announced GPT-6 and Intelligent UI on 7 October. ChatGPT can compose responses from text, visuals and interactive elements such as charts, forms and controls. OpenAI describes native components and a compiler that lets an interface appear progressively during generation. Plain text remains an option.

For a designer, format is already an argument. A side-by-side comparison invites a choice. A slider implies that something varies along a meaningful scale. A checkbox makes a messy decision look binary. When the model selects these forms, it helps frame the problem before the reader has touched anything.

That can make a conversation more useful. It also raises the standard for the answer: the control needs a sensible range, a visible unit and an understandable consequence. A beautiful slider that hides its assumptions can make a weak estimate feel unusually convincing.

Sources: OpenAI [1]

What the independent teardown adds

On 8 October, Rabi Shanker Guha published an independent reverse-engineering analysis. It describes DIL, a format combining prose, component markup and state logic; server compilation; and a sandboxed client runtime. In the observed implementation, some controls can update locally without another model call. These are the author’s findings, not a published OpenAI specification or a guarantee about every interaction.

The distinction matters to the experience. Moving a control and asking the model a new question are different kinds of work. A responsive control can help someone explore a relationship repeatedly, while the explanation around it remains available. The editorial opportunity is to make that relationship clear enough to inspect.

Sources: Rabi Shanker Guha / @rabi_guha [3]

A Pro subscription is not the Pro reasoning setting

The launch schedule began with Plus, Pro, Business and Enterprise on 7 October, with Free and Go expansion starting on 8 October. OpenAI’s release notes describe this as a Chat update; Work and Codex were unchanged by this launch. A rollout date does not establish access in a particular account.

The Intelligent UI FAQ lists support from Instant through Extra High. The Pro reasoning setting uses Astra and does not support Intelligent UI. Voice is also excluded. Some component state can survive a refresh, but does not carry across separate threads. Those boundaries matter if an interactive answer becomes part of a working brief.

Check the current web model picker. OpenAI’s model-access guide says older macOS and Windows desktop apps do not support the feature. Its Free/Go model description also differed from the launch material when checked on 9 October, so we do not treat the documentation as proof that every account has already received the same model.

Sources: OpenAI Help Center [6] · OpenAI Help Center [4] · OpenAI Help Center [5]

Give the interface a decision worth exploring

For a first creative experiment, I would use the two treatments from the opening and provide the assumptions myself. Keep the locations, available hours and required scenes explicit. Ask for a comparison that separates supplied facts from estimates, then let only the relevant quantities change. This is a proposed exercise, not a tested workflow.

Then deliberately choose an awkward value. Reduce the available time below what the required scenes demand. Does the answer expose the conflict, or quietly make the plan look feasible? A useful interface should help the director see what must be sacrificed. It should not turn every movement of a slider into a reassuring green result.

Finally, read the result without touching the controls. Can a collaborator understand the selected scenario, its units and its unresolved questions? Interaction is valuable when it leaves a decision easier to explain to another person. The interface is doing its best work when it makes the reasoning visible.