ProBackend
News & Analysis
5 days ago5 min read

OpenAI Canvas Brings Drafts and Code Into ChatGPT’s Workspace

OpenAI Canvas Brings Drafts and Code Into ChatGPT’s Workspace

OpenAI’s October 2024 Canvas launch added a side-by-side workspace for revising writing and code with ChatGPT, including inline edits, shortcuts and version restoration.

OpenAI’s Canvas changes the shape of a ChatGPT writing or coding session: instead of treating every revision as another message in a long thread, it puts the working document beside the conversation. OpenAI announced the feature on October 3, 2024, describing it as a new way to work with ChatGPT on writing and code. The practical distinction is simple, and useful: chat remains available for instructions while the draft or code stays visible as the thing being edited.

Canvas puts work beside the conversation

OpenAI said Canvas was designed for projects that need editing and revision, a weakness of the ordinary chat interface. The user and ChatGPT can work on an artifact side by side. A person can highlight a particular passage or code section to focus the model, while the conversation supplies the wider goal and context. ChatGPT can then offer inline feedback and suggestions rather than responding only with a fresh block of text in the thread.

That arrangement matters because it makes the object of the discussion easier to keep track of. The draft doesn’t disappear into a sequence of replies. OpenAI’s launch post says users can directly edit text or code in the workspace, and restore earlier versions with Canvas’s back button. It presents the feature as collaborative editing, not merely a new chat theme.

Canvas may open automatically when ChatGPT detects a task where the workspace could help. Users can also ask for it explicitly by including “use canvas” in a prompt, including when bringing an existing project into the workspace. At launch, OpenAI said Canvas was in beta and could also be selected manually from the model picker.

Writing and coding features at launch

Canvas includes task-specific shortcuts. For writing, OpenAI listed inline suggestions, controls to make a document shorter or longer, reading-level adjustment, a final polish for grammar and clarity, and an option to add relevant emojis. These are direct editing actions: the user can request a focused change and inspect the result in the same working space.

The coding shortcuts are similarly concrete. Users can ask ChatGPT to review code with inline suggestions, add log statements to help with debugging, add explanatory comments, or fix bugs by rewriting problematic code. OpenAI also listed porting code to JavaScript, TypeScript, Python, Java, C++ or PHP. Its announcement frames this as a response to a familiar problem: code revisions can be hard to follow when they are scattered across chat messages. Canvas is intended to make the edits easier to see and understand.

These controls suggest a practical way to use AI in coding without turning the interaction into a vague request for “better code.” Keep the task narrow: highlight the function that needs attention, ask for a review or a bug fix, then inspect the resulting change in the workspace. For a learning-oriented pass, ask for comments or logs and compare them with the code you already have. That guidance follows the functions OpenAI described; it isn’t a promise that the model’s changes are correct.

A side-by-side editing loop

A Canvas session can begin with a writing or coding prompt that ChatGPT recognizes as a project, or with an explicit request to use Canvas. Once the workspace opens, the user can select the relevant passage, state the desired change in chat, and review the inline response. Direct editing means the user needn’t accept every suggestion as-is. For a longer document, the length and reading-level shortcuts offer specific revision targets; for code, review, comments, logs and bug-fixing actions give the conversation a more defined job.

The ability to restore previous versions is part of that loop. It gives users a way back when a revision moves in the wrong direction, according to OpenAI’s description. Canvas doesn’t remove the need to read the result. It makes the sequence of changes more visible and gives the user a place to make or reject edits while keeping the conversation nearby.

What OpenAI reported about Canvas

OpenAI’s launch announcement also described its work on teaching GPT-4o when and how to use the workspace. The company said it used more than 20 automated internal evaluations and synthetic data techniques, including outputs distilled from o1-preview, to train core behaviors. Those included opening Canvas for writing and coding, making targeted edits, rewriting documents and providing inline critique.

OpenAI reported that, for writing tasks, its trigger behavior reached 83% on its evaluation compared with a zero-shot GPT-4o baseline given prompted instructions. The company cautioned that baseline performance is sensitive to the prompt used. It also said it intentionally biased coding behavior against opening Canvas too often, to avoid disrupting power users. These are company-reported internal evaluation results, not independent evidence of real-world accuracy or a guarantee that Canvas will choose the right mode in every session.

Availability and a measured comparison

At launch, OpenAI said Canvas was rolling out in beta to ChatGPT Plus and Team users globally, with Enterprise and Edu access planned for the following week. The company said it intended to make the feature available to Free users after beta. That describes the October 2024 rollout, not necessarily present-day availability.

Canvas’s central idea is an editable workspace attached to a chat, rather than a separate answer-only exchange. OpenAI’s announcement focuses on writing and code, inline collaboration, direct editing and version restoration. The value is less about a dramatic new coding capability than about keeping the work and the discussion in view together. That’s a modest interface change, but for iterative tasks it addresses a real friction point.

canvas puts work beside the conversationcanvas puts work beside the conversation