GPT-6 Intelligent UI guide

GPT-6 Intelligent UI lets ChatGPT answer with interactive charts, buttons, forms and small tools inside the Chat conversation. To try it, open ChatGPT on the web, select GPT-6 when available, choose a supported thinking level, and ask for an outcome with editable inputs. A special activation phrase is unnecessary. Pro thinking uses Astra and currently does not support Intelligent UI, even if you have a Pro subscription.

The useful question is what the controls help you decide. A calculator should let you change assumptions and inspect the arithmetic; a comparison should keep its source data visible. A decorative chart adds little when a sentence answers the question.

What is GPT-6 Intelligent UI?

In the October 7 announcement, OpenAI describes a library of native components and a compiler that processes the interface while the model generates it. GPT-6 chooses how text, visuals and interactions fit the question. A response can appear progressively while the model continues thinking, and ordinary text remains a valid format.

This is an interface capability in Chat. The announcement explicitly leaves the models powering Work and Codex unchanged. An attractive result is still generated output: verify the numbers, labels and behavior before using it.

Who gets access and where do you enable it?

These are announced rollout starts, not guarantees that every account received the feature on that date.

The launch page and model Help Center page are not fully synchronized: the latter still describes GPT-5.6 Luna in parts of its Free/Go guidance. Use the dates as a rollout plan and check your actual account. Do not infer current access from a subscription name alone.

For a first attempt:

1. Open ChatGPT on the web and start a Chat conversation.
2. Look for GPT-6 in the model menu. The menu label is GPT-6, rather than Sol; an explicitly selected older model stays on that model.
3. Choose Instant, Medium, High or Extra High, where your plan offers it. Pro effort is a different option and does not support this UI.
4. Describe the task, the inputs you want to change, and what should update. No special activation prompt is required.
5. Let the response finish, then change one input and check the result.

The Help Center supports the web and supported updated apps. Update mobile apps; the older macOS and Windows desktop apps do not support these capabilities. In a managed workspace, model permissions also matter. Buying a different plan should not be your first troubleshooting step.

How to write a useful Intelligent UI prompt

Specify the decision, the source data, the controls, the fixed rules and a checkable result. “Make it interactive” leaves all five open. Start with one small task and ask for a text summary alongside the controls so you can compare both representations.

Prompt 1: an API budget calculator with known answers

This example uses fictional USD rates, not OpenAI or OmniaKey prices. Its purpose is to check a generated calculator. Cache discounts, tools, retries, taxes and additional reasoning-token charges are excluded.

Create an interactive API budget calculator inside this conversation.
Use fictional rates of USD 1 per million input tokens and USD 4 per
million output tokens. Start with 5000 input tokens and 1200 output
tokens per request, 1000 requests per day, and 30 days.

Make token counts, rates, requests per day and days editable.
Per-request cost = (input tokens × input rate + output tokens ×
output rate) / 1000000. Daily cost = per-request cost × requests
per day. Period cost = daily cost × days.

Show the formulas and all three costs. Reject blank, negative or
non-numeric values; token counts, requests and days must be whole
numbers. Zero requests must cost zero. Add a reset control and a
plain-text summary. Label the rates fictional. Do not fetch prices.
If interactive controls are unavailable, say so and provide a table.

The independent calculation is (5000 × 1 + 1200 × 4) / 1000000 = USD 0.0098 per request. Keep enough decimal places to show this value; rounding it to two decimals too early changes the totals.

If one number is wrong, report the exact inputs, expected result and displayed result. Ask for a correction that preserves the other assumptions, then repeat the checks. These are acceptance cases for your generated tool, not a benchmark of GPT-6 accuracy.

Prompt 2: compare products without inventing missing details

Compare the three products in the table I paste next. Use only that
table. Show price, weight and warranty in aligned comparison cards.
Let me sort by price or weight. Keep missing values marked Unknown
and place them after known values when sorting. Preserve each source
link. Do not infer specifications or silently convert currencies.
Also provide a plain-text table with the same values.
If currencies or weight units differ, ask for conversion rules before ranking them.

Supply your data after the prompt. Check one missing warranty and one sorting change: does “Unknown” stay unknown, and does sorting preserve each product's price? Controls should rearrange evidence, not manufacture it.

Prompt 3: explore probability with an exact reference

Explain the sum of two fair six-sided dice with an interactive diagram.
Let me select a target sum from 2 through 12. Show the matching ordered
pairs, their count, and the exact probability out of 36 equally likely
outcomes. Distinguish (1, 6) from (6, 1). Use enumeration, not random
simulation. Include the full probability table and keyboard controls.

For a sum of 7, there must be 6 of 36 ordered outcomes, or 1/6. For 2, there is 1 of 36. Changing the selected sum must update the highlighted pairs and the probability together. This gives you a factual check as well as a visual one.

What are the limits?

Intelligent UI has no separate usage quota, according to the Help Center. Existing model, tool and plan allowances still apply. That does not mean unlimited reasoning, uploads or other tools, and it does not establish a universal number of interactive answers per day. Use the allowance and reset information shown in your account.

There are also practical limits. A clickable result can contain bad arithmetic, stale labels or unsupported source claims. Before keeping it, change an input, test zero and invalid values, use the keyboard, and compare the text summary with the visual output. After a follow-up revision, repeat those checks; a corrected chart can still leave an old summary behind.

Why is Intelligent UI missing or only returning text?

Check Chat → GPT-6 → a supported thinking level → web or updated supported app → workspace permissions → rollout. A missing model or unsupported Pro/Astra setting is an access issue. A plain-text answer with supported access may simply be the format ChatGPT chose. Request the useful controls explicitly once; a stronger prompt cannot grant account access.

Can you turn off the visuals?

On the web, open Settings and turn off Layout and visuals. OpenAI says this reduces visual elements but does not guarantee a fully classic or text-only experience. You can also request a plain-text answer for the current task; treat that as a formatting request, not a different model or allowance.

Intelligent UI, Visualize, Sites or an API?

Choose the surface based on where the result needs to live.

OpenAI documents Visualize and Sites separately. You do not need to treat the Visualize workflow as the activation procedure for this Chat release.

For developers, the October 7 API changelog updates chat-latest, but a model snapshot is not ChatGPT's native interface renderer. Adding UI to an MCP server is a documented integration route that uses UI resources and an iframe. It requires application work; switching an API model name does not supply that UI.

If you are building your own tool, compare models and current routes in the OmniaKey model catalog. Our GPT-6 Sol vs Luna guide covers model selection, while the GPT-6.1 Sol review covers coding and API work. Keep those decisions separate from access to ChatGPT's interface.