Until now, the division of roles was clear: a language model delivers text, and the product team decides how that text is presented — as a chat bubble, as a table, as a summary in a dashboard. With the release of GPT-6, OpenAI shifts this boundary. Under the name Intelligent UI, the model no longer outputs its answers only as running text, but as an interactive interface: with diagrams, buttons and forms.
Technically speaking, this is a small step with major consequences for product design. In future, the question will no longer be solely what a model answers, but also in what form it answers — and who controls that form.
What OpenAI has announced
The key points of the announcement can be summarised briefly:
- Intelligent UI: Answers are output as interactive interfaces that can contain diagrams, buttons and forms.
- Parallel thinking and answering: The system is intended to process and output simultaneously and thereby respond noticeably faster, without the answer quality suffering as a result — according to OpenAI, it even improves.
- Two variants: Paying customers receive the stronger variant GPT-6 Sol, free accounts work with GPT-6 Luna.
More details on the technical implementation are not currently public. The extent to which the generated interfaces can be influenced in terms of design, whether they can be bound to a company's own design system and how they can be used via interfaces cannot be derived from the announcement.
Why generated interfaces are something different from generated text
Text is tolerant. A paragraph that is formulated somewhat awkwardly remains readable. An interface is not. As soon as a model generates form fields and buttons, interactions arise — and with them expectations of consistency, accessibility, error handling and traceability.
A diagram also appears more authoritative than a sentence. Anyone who sees a number in a bar chart questions it less often than the same value in running text. That is not an argument against generated visualisations, but an argument for keeping the underlying data visible and verifiable.
In addition: a form generated by a model must lead somewhere. The exciting question for companies is less whether the model can draw an input field, but what happens to the data entered — and which system it ends up in.
What shifts for product teams
When a model delivers interface building blocks, the division of labour in the project changes. Three shifts are emerging:
- From layout to framework: Design work shifts from the arrangement of individual elements to the definition of rules — which components are permissible, which states exist, where the boundaries lie.
- From screen to data model: The more dynamic the presentation, the more important a clean, well-documented data basis becomes. A generated interface is only as good as what it displays.
- From testing the result to testing the behaviour: Classic acceptance checks examine finished screens. Dynamically generated interfaces instead require checks along typical scenarios.
Speed as a product feature
The second part of the announcement is easy to overlook, but carries practical weight: through parallel thinking and answering, the system is intended to respond significantly faster. Waiting time is not a side issue in interactive applications. An interface that only appears after a noticeable delay is used differently from one that is there immediately. The closer AI-supported interaction comes to real time, the more realistic use cases become in which users adjust things several times in succession.
Two model variants, two experiences
The split into GPT-6 Sol for paying accounts and GPT-6 Luna for free use is not new in itself, but deserves attention. Anyone building processes on such services should know which variant is working in the respective use case. Different model tiers can lead to different results — with generated interfaces, possibly not only in terms of content but also in the depth of presentation.
Caution when transferring to your own projects
Intelligent UI is initially a feature within ChatGPT. It does not automatically follow from this that generated interfaces can be adopted into your own websites or specialist applications at short notice. Anyone planning today should distinguish between two things: observing a trend and committing to a specific technology.
A level-headed approach is therefore sensible: identify internal use cases in which dynamic presentation brings genuine benefit — for example evaluations, configurators or guided input. And in parallel, do the homework that remains useful regardless of the provider: structured content, clean interfaces, documented data sources.
Assessment
For web and software projects, the focus shifts: it is not the pixel-perfect design of individual pages that becomes the decisive factor, but the quality of the content, data and interfaces behind them. Anyone who sets up their systems so that content is available in a structured and machine-readable form remains able to act — regardless of whether an interface is built classically or will in future be partly generated.
Decisions about specific tools, on the other hand, can be made calmly once reliable experience from productive use is available.