ChatGPT Intelligent UI: Interactive Charts, Calculators and Diagrams

Check your plan, mode and client

Swipe wide tables horizontally on a phone.

Model and client eligibility · October 7 rollout announcement.

Source discrepancy: the release notes and UI help say Free/Go rollout uses GPT-6 Luna; the model-availability article still lists GPT-5.6 Luna. We cannot resolve that into an account-wide availability claim. The model menu and working controls in your account are the practical check.

Start with a decision, then request an interaction

1. Name the question and supply the inputs you are allowed to share. Separate known values from estimates.
2. Describe what should change: a quantity field, a date filter, a scenario switch or a diagram node. Specify the dependent output.
3. Ask for the formula or reasoning in readable text beside the interface. A control with no explanation is difficult to audit.
4. Change one input at a time. Check that the relevant result changes and an unrelated result stays stable.
5. Keep a plain-text or table fallback for someone who cannot open the interactive view.

ChatGPT chooses a layout automatically; an explicit request guides it but does not guarantee a particular widget. Some state can remain in the same thread, not across threads. Layout and state behavior.

Four original questions and prompt briefs

These fictional briefs are starting points, not screenshots or successful runs. All prices and quantities below are invented inputs, not product quotes.

1. Which subscription is cheaper at my volume?

Compare fictional Plan A ($12/month plus $0.08 per item) with Plan B ($30/month including 300 items, then $0.04 per extra item). Make an item-count input from 0 to 1,000 and show both monthly totals. Show the piecewise formulas and mark the cheaper plan. Do not invent taxes or discounts. Include a plain table at 0, 100, 300 and 500 items.

At 100 items, the checks are $20 and $30. At 500, they are $52 and $38. Inspect the boundary at 300 and reject negative counts. A pretty comparison with a wrong included allowance is not usable.

2. How much should each traveler pay?

Create an editable bill splitter for four fictional travelers: lodging $640, transport $120 and shared food $96. Split all three costs equally. Allow me to exclude one person from one cost and show each person's amount and the total. Show the rounding rule and reconcile the split to the original bill.

The initial check is $856 total and $214 per person. Excluding someone from lodging should redistribute only lodging; transport and food must stay equally split.

3. Is the apparent growth just a partial month?

Use this fictional dataset: July 120 signups, August 150, September 135, October 60 through October 10. Build a chart with a switch to exclude incomplete months. Label October as partial, show the underlying table and do not forecast the rest of October.

Verify that changing the switch does not change the three completed months. A percentage computed against partial October should carry that limitation, not a confident decline narrative.

4. Where does an approval delay occur?

Draw an interactive diagram for draft → reviewer → approved → published. Add a rejected branch back to draft. Clicking each node should reveal its entry rule, exit rule and responsible role. Do not connect to any system or actually publish anything.

Follow the rejection branch and check that publication cannot precede approval. The diagram describes a process; it does not enforce it in your tools.

Validate the answer separately from the interface

Keep a small reference table outside the generated view. Check one ordinary case, a zero case and a boundary case with an independent calculator or spreadsheet. For dates, specify time zone and whether the end date is included. For money, separate currency, tax and rounding. For percentages, name the denominator.

Inspect whether filters change the calculation or merely hide rows. If a chart summarizes uploaded data, compare its row count and totals against the original file. Missing data must remain missing rather than becoming zero. Do not share a result until the source, assumptions and unresolved differences are visible.

For consequential decisions, use the relevant qualified review and your source system. Interactivity is a presentation feature, not proof that a number or recommendation is correct.

Why did I only get text?

Swipe wide tables horizontally on a phone.

UI preferences · Usage and client limitations.

When to use it, and when to choose another route

Good fit: exploring a few scenarios, explaining a process, or discussing a small dataset with a colleague who can verify the inputs. Poor fit: a production dashboard requiring audited refreshes, a public calculator with uptime obligations, or a workflow that must enforce approvals.

A widget inside a conversation does not establish hosting, an accessible public URL, authentication, data storage or a deployment licence. If your outcome is a website, define those requirements separately. For the broader product and subscription decision, read our ChatGPT profile. For connected business metrics and query review, see the Claude Dashboards guide.

Next step: pick one fictional brief above, replace only non-sensitive inputs, and verify two known outputs before relying on the view.