ChatGPT Canvas & Code Interpreter

Canvas is a side-by-side editing surface. Instead of regenerating a whole document for every tweak, ChatGPT opens a panel you and the model edit together — conversation on the left, living document on the right.

Two ways to open it:

• Automatically — when ChatGPT generates content longer than roughly ten lines, or detects a writing or coding task worth editing.
• On request — add "use canvas" to your prompt.

Try this in a fresh chat:

> Use canvas to draft a one-page onboarding email for new customers
> of a project-management app. Friendly but professional, with a
> short welcome, three things to do first, and a sign-off.

A panel slides in with the draft — editable text, not a frozen reply. Click in and type to change it yourself.

The real power of Canvas is targeting one part of the document instead of the whole thing. Highlight a sentence or paragraph and ChatGPT focuses only on your selection.

Highlight the "three things to do first" list, then type:

> Make these three steps more specific and start each one with a verb.

Only that section changes; the rest is untouched.

Canvas also exposes one-click shortcuts as panel buttons:

Code Interpreter (also labelled data analysis) lets ChatGPT write and run real Python in a sandbox to crunch files you upload. You ask in plain English; it writes the code, runs it, reads any errors, and fixes them before answering.

Start a new chat, click the attachment control next to the message box, upload your CSV, and ask:

> Here is a CSV of monthly sales. Profile it: how many rows, what
> columns and types, any missing values, and the date range. Then
> tell me the three highest-revenue months and the month-over-month
> growth rate for the last six months, in a table.

ChatGPT loads the file, inspects it, and replies with a profile, a scrollable interactive table, and your analysis. The file and prior results stay available for the whole chat, so you can keep refining.

Now ask for visuals and a file to keep. Code Interpreter builds charts and tables from your file, then hands you a download link — the sandbox lets you upload files and download the results.

> Plot total revenue by month as a line chart, labelling the axes.
> Then add a month-over-month growth column and export the full
> table as an Excel file I can download.

Then verify before you trust:

> Show me the exact rows and formula you used to calculate that 12%
> growth figure, so I can check it myself.

The third built-in tool is web search. When ChatGPT searches, it returns timely answers with inline citations — hover to preview a source, click to open it. Search is available across the tiers, including Free, Plus, Team, Edu, and Enterprise.

It often searches on its own for time-sensitive questions, but you can force it with the web-search control near the message box, or just ask:

> Search the web. What are the current pricing tiers for ChatGPT,
> Claude, and Gemini? Give me a comparison table and cite the
> official pricing page for each one.

Read the answer, then click the citations. The model summarises sources; the sources are where the truth lives. For anything fast-moving — prices, models, limits — open the linked page before acting.

These three tools overlap just enough to confuse people. The mental model:

They also combine — a realistic workflow chains all three:

> Search the web for the latest electric-vehicle sales figures and
> cite the sources. Then chart year-over-year growth from those
> numbers, and use canvas to draft a short briefing memo.

One request researches with citations, analyses with code, and drafts in an editable surface — research to deliverable, one chat.