OpenAI Introduces Intelligent UI: Visual Answers You Can Click, Adjust, and Play With

Ask ChatGPT how to find the North Star, and it can give you a detailed explanation. Ask it to teach you mahjong, and it can list the rules. But after reading the text, you still have to match the directions to the sky and map “four sets and a pair” onto the tiles in front of you. AI has already explained it once; you still have to piece it together in your head.

Intelligent UI, introduced with GPT-6, changes that step. When you look for a star, the answer becomes a star chart you can explore one step at a time. When you learn the rules, the tiles sit alongside the explanation. When you need a small tool, an interactive calculator or game can appear in the chat. You can click, adjust, and experiment directly within the answer, connecting seeing, understanding, and using it.

What is Intelligent UI?

Intelligent UI is the intelligent answer interface introduced in ChatGPT with GPT-6. Depending on your question, it combines text, graphics, and interactive controls into an answer suited to the task, so you can view, adjust, and use it within the chat.

Its main features:

• Choose an answer format to fit the question: Show comparisons side by side, explain a process through diagrams or steps, and keep simple questions as straightforward text answers.
• Use graphics and steps to aid understanding: Put component relationships, directions on a star chart, tile rules, or painting details in front of you, reducing the effort of constructing the picture in your head.
• Let you explore changes yourself: Use step buttons, forms, or other controls to follow a process, change the input, and observe the results.
• Use small tools directly in the chat: Generate calculators, bill splitters, or small games on request, ready to operate as soon as the answer appears.
• Adjust an existing answer directly: Select conditions such as color or style in the interface and see the results of different choices.
• Display the interface as it is generated: The interface appears progressively with the answer; you do not have to wait for all the content before you start viewing it.

Why are some everyday tasks hard to explain with text alone?

Room color schemes, travel plans, origami instructions, and bicycle explanations can all run to many paragraphs and still leave you unsure: What do the colors look like together? How do you get between attractions? Which line should you fold along? Where exactly is a particular part installed?

Tasks involving spatial position, structural hierarchies, or hands-on steps are often difficult to convey intuitively through continuous text alone. Whether you are learning how mechanical parts work together, planning stops along a route, or folding paper along successive creases, you have to work to turn words into spatial relationships and a sequence of actions.

Intelligent UI puts those relationships directly into the answer. Bicycle parts can be viewed by system; travel routes appear on a map with photos of the attractions; room color schemes use swatches for comparison; and origami uses illustrated steps. You can match the explanation to what you are looking at without relying entirely on words to imagine locations and actions.

Graphics and controls still need to work with text. A patio planting plan combines a diagram of container placement with written explanations, while a simple question can receive a direct text answer. The point is to choose side-by-side comparisons, diagrams, interactive components, or concise text according to what helps people understand and use the answer, rather than making every response complicated.

How answers become easier to understand and use

1. Follow a star chart to find the North Star

• User question: “How do I find the North Star?” (How do I find the North Star?)
• Generated interface: Instead of a barrage of directional terminology, the screen shows an interactive night-sky chart with 4 sequential step buttons at the bottom:
  1. Find the Big Dipper (Find the Dipper);
  2. Identify the two pointer stars (Find the pointers): Highlight Merak and Dubhe at the outer edge of the bowl;
  3. Extend the line (Follow the line): Draw a dashed line labeled “5x distance”;
  4. Reveal Polaris (Reveal Polaris): Highlight the target on the chart and display the completion message, “You found the North Star!”
• Practical value: Abstract spatial instructions become a “pocket planetarium” you can work through step by step at your own pace.

2. Match mahjong rules to actual tiles

• User question: “Teach me how to play mahjong.” (Teach me how to play mahjong)
• Generated interface: The interface lays out an illustrative hand: one through three of Characters, one through three of Circles, one through three of Bamboo, three East Winds, and two Red Dragons. The first three groups of one, two, and three are sequences; the three East Winds form a triplet; and the two Red Dragons form a pair. It prominently labels the basic winning-hand formula, “The objective: four sets and a pair” (The objective: four sets and a pair). Cards below separately introduce the Characters, Circles, and Bamboo suits and explain that each numbered tile has 4 identical copies.
• Practical value: Beginners can match each rule to the actual tiles before having to memorize unfamiliar terms for sequences, triplets, and the pair.

3. Play a retro handheld game directly in the chat

• User question: “Make a retro gaming console I can play.” (Make a retro gaming console I can play)
• Generated interface: ChatGPT renders a retro handheld console directly in the conversation, complete with a classic green LCD screen. A paddle game runs on the screen, showing the live score for the player (YOU) and ChatGPT (CHAT). A clickable virtual directional pad and A/B buttons sit below the screen, letting you control the game directly within the chat.
• Practical value: You can use small interactive tools without leaving the conversation. Practical tools such as a bill splitter for a group meal (Bill splitter) or a retirement savings calculator (Retirement savings calculator) also let you enter numbers and run calculations on the spot.

4. Look at the painting to understand why Monet’s water seems to move

• User question: “How did Monet make water look like it’s moving?” (How did Monet make water look like it's moving?)
• Generated interface: The answer displays the full view of Monet’s Water Lilies, then breaks it down into four close-up detail cards:
  • Distorted reflections (Distorted reflections): Reflections of trees and bridges stretch downward, then break into patches of color. The brain reads these distorted shapes as the movement of water;
  • Broken brushstrokes (Broken brushstrokes): Short horizontal strokes interrupt reflections to suggest ripples. Instead of one continuous field of blue, the surface consists of dozens of small brush marks;
  • Contrasting colors (Contrasting colors): Blues, purples, greens, and a few warm colors sit directly beside each other rather than blending smoothly, creating a shimmer like sunlight dancing across water;
  • Layers of depth (Layers of depth): Lily pads float on the surface while reflections of clouds and trees appear below, moving your gaze between the surface, reflections, and a sense of depth.
• Practical value: An explanation of painting techniques becomes a side-by-side analysis of text and visual detail. On-screen explanations and close-ups reinforce each other, helping you observe how the painting creates a sense of moving water.

5. Adjust colors and styles on a moodboard

• User question: “My dog needs a cool vest. Can you make a moodboard?” (My dog needs a cool vest. Can you make a moodboard?)
• Generated interface: The interface generates a “dog vest moodboard” with a color palette and style tags, offering options such as retro sporty (Retro sporty), streetwear (Streetwear), maximalist (Maximalist), soft and playful (Soft and playful), and outdoorsy (Outdoorsy). Clicking different tags immediately updates the dog reference images and clothing looks on the screen.
• Practical value: Rather than repeatedly typing requests for changes, you can select options in the generated interface to explore preferences and make adjustments directly.

6. Break a bicycle down into parts you can browse

• User question: “Break down the design of a 7-speed bike.” (break down the design of a 7‑speed bike)
• Generated interface: Six filter buttons appear above the answer: All (All), Frame (Frame), Wheels (Wheels), Drivetrain (Drivetrain), Brakes (Brakes), and Cockpit (Cockpit). Below them are corresponding structural explanations, including how the double-triangle frame carries weight and the fork steers; how pedaling drives the chain and rear wheel across 7 gears; and how handbrake cables pull brake pads against the wheel rims to slow the bicycle.
• Practical value: Buttons let you switch between subsystems as needed, breaking a complex industrial product into categorized modules of images and text.

7. Adjust the guest count and take a shopping list with you

Preparing a meal for friends involves more than choosing a menu: you need to calculate how much food to buy and decide when to start cooking. In the website’s comparison using the same group-meal question, GPT-5.6 Instant provides a menu, a shopping-quantity table for fixed guest counts, and a schedule. GPT-6 Instant turns the guest count into an adjustable control, calculates quantities according to the number of guests, and provides a “Copy shopping list” button. The answer also includes a checkable cooking timeline, so you can mark preparation tasks as complete.

The value here is bringing the conditions, calculated results, and actions together. If the guest count changes, you can adjust it, check the quantities, and copy the list to take shopping. Once preparation starts, follow the timeline and check off the steps. You get more than a suggestion: you can adjust the plan, take the results with you, and track execution within the same answer.

8. Change the input to understand complex concepts

Some questions are difficult because it is hard to see why the result changes when a condition changes. Intelligent UI can explain through visuals and interaction, letting you change inputs and observe results or understand a problem step by step. Learning scenarios listed on the website include the central limit theorem (a statistical theorem explaining why standardized sample means approach a bell-shaped distribution under certain conditions), GDP, drone photography, and the Monty Hall problem (a probability problem asking whether you should switch after choosing a door and the host opens another empty door). Such answers put written explanations alongside observable changes to help you understand the relationship between conditions and results.

Behind the scenes: how does the model build an interface as you chat?

Producing a passage of text is straightforward for a large language model. Building a well-structured, smoothly interactive interface within a conversation mainly relies on three technical foundations:

The process can be summarized as follows: the user asks for a specific task → GPT-6 determines the most suitable format for the question (plain text, a chart, or interactive components) → it arranges elements from a native streamable component library → a compiler processes them in real time, allowing the interface to appear progressively as it is generated.

1. Native streamable component library (Native streamable component library): OpenAI built a native component library that works on the web and mobile, covering common elements such as buttons, forms, and charts. The library provides a consistent design foundation for each answer while letting the model decide flexibly how to combine components.
2. Streaming compiler (Streaming Compiler): A dedicated streaming compiler processes the interface in real time while the model generates it. Users therefore do not have to wait for the entire answer: the interface appears progressively as generation proceeds.
3. Design judgment training (Design judgment training): The development team expanded its training methods to assess the clarity, usefulness, and completeness of generated interfaces. The model learns to organize information, decide when to use interactive components, and judge when a brief plain-text answer is more appropriate.

OpenAI also says the model’s interface design judgment still needs improvement, and the range of things it can create is expanding. Choosing a suitable interface for different questions remains a capability under continuous development.

Faster responses: ChatGPT can answer while it thinks

Reasoning models have helped users solve harder problems, but users have generally had to wait until the model finished thinking before seeing an answer. ChatGPT can now start answering while it continues to think. OpenAI trained the models to account for users’ waiting time and build answers progressively from the knowledge and analysis available at that point.

The model learns to develop a complete answer across multiple response segments, with each segment adding useful information rather than filler, while keeping the whole answer as coherent and factual as if it had been written in one go.

The measure here is “how soon the answer starts.” It does not mean the total time to finish an answer or task also falls by 44%. According to OpenAI’s internal evaluations:

• High-value everyday Agent tasks: At the Extra High reasoning level, GPT-6 starts answering in about the same time as GPT-5.6 at Medium, while scoring better overall than GPT-5.6 Extra High.
• Questions requiring web search: GPT-6 Instant’s average wait before it starts answering is 44% shorter than GPT-5.6 Instant’s.
• Retrieval and answer quality: The model makes better judgments about when to look things up and more reliably finds information that supports its answers. In internal evaluations of difficult questions, it accurately addresses the core of the user’s question more often than GPT-5.6.

Stronger safeguards and defenses across multiple turns

GPT-6 incorporates advances in defenses from the Astra series, with a focus on more accurately recognizing real-world risks:

• Resisting multi-turn adaptive jailbreak attempts (Multi-turn adaptive jailbreak attempts): GPT-6 can recognize safety-relevant context across turns when attackers gradually probe through a conversation and attempt to bypass safety policies.
• Curbing misuse in high-risk domains: It strengthens safeguards around high-risk scenarios such as cyberattacks, biological hazards, and violent crime.
• Reducing blanket refusals: It improves judgments about safety boundaries to avoid unnecessary refusals of harmless questions that contain sensitive terminology.
• Recognizing its own limitations: When it lacks the information or tools needed to answer, the model is better at recognizing and explaining its limitations.

Rollout and model differences

Models and rollout arrangements differ by user group:

This update applies only to ChatGPT’s chat interface (Chat tab). The underlying models powering Work and Codex remain unchanged in this release.

The next step in interaction: software adapts to people

For decades, people have learned to use predefined software, moving between fixed interfaces designed for particular tasks.

OpenAI sees a different future: software adapts to people instead of people adapting to software. Intelligent interfaces are a step in that direction. As the model’s design judgment improves, users could simply describe the task they want to accomplish and have the system build an interface suited to it, making information easier to obtain and work with.

Download the complete bilingual video compilation · Download all six videos and editable subtitles