ChatGPT gets a workspace called Canvas

Anyone who has used ChatGPT to write a long document or refactor code knows the pain: you ask for a tweak in paragraph three, and it dutifully reprints all 2,000 words from scratch.

OpenAI just launched Canvas, a brand-new interface built for writing and coding projects that push past the limits of a standard chat box. Marking the first major visual redesign of ChatGPT in two years, the feature rolls out in beta today for Plus and Team subscribers worldwide, with Enterprise and Edu access coming next week and free users getting it once it exits beta.

Moving beyond the chat box

Canvas opens up a separate workspace side-by-side with your conversation. Instead of just talking back and forth, you and ChatGPT can now collaborate directly on the same document or codebase.

Users hold full control over the canvas. You can type directly into the document, highlight specific lines to ask for targeted feedback, and hit a back button to restore previous versions.

For writers, OpenAI added handy shortcut actions. You can ask ChatGPT to suggest inline edits, adjust text length, add final polish for clarity and grammar, add emojis, or even change the reading level across a scale from Kindergarten to Graduate School.

For developers, the feature takes inspiration from modern code review workflows. Shortcuts allow you to review code, insert logs for debugging, add explanatory comments, fix bugs automatically, or translate code into JavaScript, TypeScript, Python, Java, C++, or PHP.

Built on o1 synthetic training

Under the hood, OpenAI built Canvas using GPT-4o, but getting the AI to act like a thoughtful collaborator required new training techniques. The team used synthetic data distilled from OpenAI o1-preview to train the model on when to open the workspace, when to suggest comments, and when to rewrite versus edit targeted lines.

Triggering the interface correctly was one of the biggest hurdles. OpenAI taught the model to open Canvas for tasks like drafting a full blog post, while intentionally keeping it closed for quick Q&A prompts like recipe recommendations. In writing tasks, the trigger accuracy reached 83%, and in coding tasks it hit 94%—with code triggers intentionally tuned to be conservative so power users aren't interrupted.

When editing text, Canvas outperformed a prompted GPT-4o baseline by 18% on targeted modifications. In human evaluations, its inline comments showed a 30% jump in accuracy and a 16% boost in overall quality.

Why it matters

We think Canvas is a overdue, highly necessary evolution for ChatGPT. The single-column chat interface was great for casual Q&A, but working on complex text or debugging code in a vertical stream has felt clunky for months.

Anthropic made waves earlier this year with Claude Artifacts, proving that users desperately want a persistent workspace next to their chat. Canvas goes a step further by making that canvas natively interactive and editable, complete with dedicated review controls.

For everyday users and coders, this is worth diving into right away. The friction of copying snippets back and forth between ChatGPT and Google Docs or VS Code drops dramatically. OpenAI's challenge will be keeping those automated triggers accurate without annoying power users, but this marks a clear shift: AI assistants are leaving the chatbot phase behind to become real-time co-editors.