What is canvas in ChatGPT, and how does it change writing and coding with AI?

Canvas is a new workspace in ChatGPT, launched in beta on October 3, 2024, that opens your document or code in a separate window beside the chat so you and ChatGPT can edit it together. You can highlight a passage to get targeted changes, edit directly yourself, use one-click shortcuts such as "adjust the length" or "fix bugs," and go back to earlier versions.

What is canvas in ChatGPT, and how does it change writing and…

Yesterday, Thursday, October 3, 2024, OpenAI introduced canvas, a fresh way to write and code together with ChatGPT. Instead of every change arriving as a fresh reply further down the chat, your draft sits in its own window next to the conversation, where you and ChatGPT can both edit it.

OpenAI calls it the first major change to ChatGPT's visual interface since the service launched two years ago. It is an early beta, rolling out to Plus and Team users around the world from launch day, with Enterprise and Edu users next week. OpenAI plans to offer it to free users once the beta ends.

What canvas is

In a normal chat, revising a document is clumsy. You ask for a change, ChatGPT rewrites the whole thing, and you scroll to compare versions. If one sentence is wrong, you often have to describe exactly where it is.

Canvas fixes that with a shared workspace, much like working on a document with an editor sitting beside you:

Highlight to focus. Select a sentence, paragraph or block of code, then ask for a change. ChatGPT works on that part, with the rest of the document in mind.
Edit yourself. You can type directly in the canvas, just as in a word processor or code editor.
Shortcuts. A menu offers common actions in one click.
Version history. A back button restores earlier versions of your work.

Canvas opens on its own when ChatGPT judges it would help, for example when you ask for a longer piece of writing or a coding task. You can also type "use canvas" in your prompt. During the beta, you can choose "GPT-4o with canvas" in the model picker.

The shortcuts

For writing	What it does	For coding	What it does
Suggest edits	Inline comments and suggestions	Review code	Inline suggestions to improve the code
Adjust the length	Makes the text shorter or longer	Add logs	Adds print lines so you can trace what the code does
Change reading level	Rewrites for any level, kindergarten up to graduate school	Add comments	Explains the code in plain-language notes
Add final polish	Checks grammar, clarity and consistency	Fix bugs	Finds and rewrites code that causes errors
Add emojis	Adds emojis for emphasis and color	Port to a language	Converts code to Python, JavaScript, TypeScript, Java, C++ or PHP

In a demo for TechCrunch, Daniel Levine, a product manager at OpenAI, had ChatGPT draft an email, used a slider to make it shorter or longer, highlighted a sentence to make it friendlier, and had the whole message rewritten in another language. For code, he asked for a small web server in Python, added comments with one button, and used "review code" to get suggested changes he could accept, edit or turn down.

How OpenAI built it

Canvas is not a separate model but GPT-4o trained to behave like a collaborator. OpenAI taught it five core behaviors: deciding when to open a canvas, producing different kinds of content, making targeted edits, rewriting whole documents, and giving inline critique.

The hardest part was knowing when to open a canvas at all. A request to "write a blog post about the history of coffee beans" should open one; "help me cook a new recipe for dinner" should not. OpenAI reports that the trained model opens canvas correctly for 83% of writing tasks and 94% of coding tasks in its tests, an improvement on GPT-4o simply instructed by a prompt. For code, it deliberately made the model slower to open a canvas, so experienced users are not interrupted.

A second question is when to edit a small part and when to rewrite everything. The rule OpenAI chose is easy to remember: if you select text, ChatGPT makes a targeted edit; if you do not, it tends to rewrite.

To train these behaviors, OpenAI used synthetic data, training examples created by AI rather than written by people, including answers distilled from its o1-preview reasoning model. It tracked progress with more than 20 automated tests, and used human reviewers to judge comment quality. Those reviewers found the canvas model's comments 30% more accurate and 16% better in quality than prompted GPT-4o.

Why it matters this week

Canvas moves ChatGPT from a conversation toward a workspace. That matters for anyone whose work involves revision: students drafting essays, people writing reports or cover letters, and programmers fixing code. Being able to point at the exact passage you mean removes much of the back-and-forth.

It is also part of a wider trend. TechCrunch notes that Anthropic's Artifacts, launched in June, and the coding tool Cursor offer similar side-by-side workspaces. Editable workspaces accept a simple reality: AI rarely gets a long piece right in one try, but it can give you a strong first draft that you then shape together.

Try it: Paste a paragraph you wrote recently into canvas. Highlight the weakest sentence and ask for two alternatives, then use "Change reading level" to see the paragraph at a middle-school level. Keep only the changes you agree with.

How do you open canvas and get the most out of it?

If you have ChatGPT Plus or Team, select "GPT-4o with canvas" in the model picker, or simply write "use canvas" in your request. Enterprise and Edu users get it next week; free users later, after the beta.

Start with a clear request that gives the purpose, the reader and the length:

Prompt

Use canvas. Draft a 200-word welcome email for new volunteers at a community garden. Friendly tone, plain language, and include the first meeting date as a placeholder.

Then work in small steps. Highlight one part at a time and say what you want changed, use the shortcuts for broad passes like length or polish, and make your own edits directly when you know what you want. Use the back button freely; earlier versions are not lost. As always, read the final text yourself and check any facts, names and numbers before you send or publish it.