With the release of Canvas, a new writing and coding interface from OpenAI, ChatGPT can now do more than just have conversations. With ChatGPT, users can work on writing and coding projects through the interface, which provides real-time edits, comments, and ideas. Canvas is currently available to ChatGPT Plus and Team members worldwide. Next week, it will also be made available to Enterprise and Edu users, with plans to make it available for free users after the beta phase. Canvas can be manually chosen in the model picker and is integrated with GPT-4o.  

Better collaboration with ChatGPT  

Users also leverage ChatGPT for help with writing and code. Although the chat interface is easy to use and works well for many tasks, it’s limited when you want to work on projects that require editing and revisions. OpenAI claims that Canvas offers a new interface for this kind of work.  

“With canvas, ChatGPT can better understand the context of what you’re trying to accomplish. You can highlight specific sections to indicate exactly what you want ChatGPT to focus on. Like a copy editor or code reviewer, it can give inline feedback and suggestions with the entire project in mind”, says OpenAI.  

OpenAI stated, “You control the project in canvas. You can directly edit text or code. There’s a menu of shortcuts for you to ask ChatGPT to adjust writing length, debug your code, and quickly perform other useful actions. You can also restore previous versions of your work by using the back button in canvas.”  

Canvas opens automatically when ChatGPT detects a scenario that could be helpful. Users can also include “use canvas” in their prompt to open Canvas and use it to work on an existing project.  

Writing shortcuts include:  

• Suggest edits: ChatGPT offers inline suggestions and feedback.  
• Adjust the length: Edits the document length to be shorter or longer.  
• Change reading level: Adjust the reading level from Kindergarten to Graduate School.  
• Add final polish: Check for grammar, clarity, and consistency.  
• Add emojis: Add relevant emojis for emphasis and colour.  

Coding in Canvas  

Coding is an iterative process, and it can be hard to follow all the revisions to your code in chat. Canvas makes tracking and understanding ChatGPT’s changes easier, and OpenAI claims that they plan to continue improving transparency in these kinds of edits.  

Coding shortcuts include:  

• Review code: ChatGPT provides inline suggestions for improving your code.  
• Add logs: Insert print statements to help debug and understand your code.  
• Add comments: Add comments to the code to make it easier to understand.  
• Fix bugs: Detects and rewrites problematic code to resolve errors.  
• Port to a language: Translates your code into JavaScript, TypeScript, Python, Java, C++, or PHP.  

Training the model to become a collaborator  

OpenAI stated, “We trained GPT-4o to collaborate as a creative partner. The model knows when to open a canvas, make targeted edits, and fully rewrite. It also understands broader context to provide precise feedback and suggestions”.  

To support this, the research team developed the following core behaviours:  

• Triggering the canvas for writing and coding   
• Generating diverse content types  
• Making targeted edits  
• Rewriting documents  
• Providing inline critique  

The team measured progress with over 20 automated internal evaluations. They used novel synthetic data generation techniques, such as distilling outputs from OpenAI o1-preview, to post-train the model for its core behaviours. According to the blog post, this approach allowed them to rapidly address writing quality and new user interactions, all without relying on human-generated data.  

A key challenge was defining when to trigger a canvas. “We taught the model to open a canvas for prompts like “Write a blog post about the history of coffee beans” while avoiding over-triggering for general Q&A tasks like “Help me cook a new dinner recipe.” For writing tasks, we prioritized improving “correct triggers” (at the expense of “correct non-triggers”), reaching 83% compared to a baseline zero-shot GPT-4o with prompted instructions,” the blog post read.  

It’s worth noting that the quality of such baselines is highly sensitive to the specific prompt used. With different prompts, the baseline may perform poorly but differently—for instance, by being evenly inaccurate across coding and writing tasks, resulting in a different distribution of errors and alternative forms of suboptimal performance. “For coding, we intentionally biased the model against triggering to avoid disrupting our power users. We’ll continue refining this based on user feedback,” OpenAI added.  

A second challenge involved tuning the model’s editing behaviour once the canvas was triggered—specifically deciding when to make a targeted edit versus rewriting the entire content. “We trained the model to perform targeted edits when users explicitly select text through the interface, otherwise favouring rewrites. This behaviour continues to evolve as we refine the model,” OpenAI noted.  

Finally, training the model to generate high-quality comments required careful iteration. Unlike the first two cases, which are easily adaptable to automated evaluation with thorough manual reviews, measuring quality in an automated way is particularly challenging. Therefore, the team used human evaluations to assess comment quality and accuracy. OpenAI’s integrated canvas model is claimed to outperform the zero-shot GPT-4o with prompted instructions by 30% in accuracy and 16% in quality, showing that synthetic training significantly enhances response quality and behaviour compared to zero-shot prompting with detailed instructions. 

OpenAI