OpenAI introduces ChatGPT Work with cloud and local versions

Summary

Simon Willison provided a hands-on review of ChatGPT Work, an OpenAI product that layers internet-connected code execution, a headless Chrome browser, a persistent shared file system, and sub-agent orchestration on top of the standard chat experience within the paid tier.

Why it matters

This marks a concrete shift from stateless, isolated chats to agentic workflows with real web access and cross-session memory. These capabilities previously belonged to developer tools like Codex. Placing these features behind a $20/month paywall, rather than offering them in the free or $8 tiers, suggests OpenAI is positioning autonomy as part of its monetization strategy. Specifically, open internet code isolation narrows the gap between a chatbot and a general-purpose computing agent.

Highlights

• Code execution with unlimited internet access (repo cloning, pip/npm installations, API calls), which is the most distinct differentiator when compared to both standard chat and Claude's permission-listed container.
• Headless Chrome with Playwright-style JS evaluation and 2FA transfer flows, enabling non-developers to scrape and interact with live sites directly from their phones.
• A persistent /workspace file system shared across concurrent work sessions (Simon already has 171 draft folders), enabling stateful, multi-step projects that per-session chat isolation cannot support.

Critical take

Willison describes the product as 'extraordinarily confusing' and labels OpenAI's official chat-vs-work comparison guide as 'almost entirely useless.' This constitutes a sharp critique of the product's positioning. The article also cuts off right in the middle of the security discussion. Open internet access, a fully-featured browser, and a shared writable file system create a significant attack surface. It is a notable omission that a comprehensive review fails to provide a concrete answer to the question: 'Is this safe?'