OpenAI Operator: ChatGPT is Free to Roam the Internet and (Mostly) Does What You Say

OpenAI Operator turns ChatGPT into a web-running agent: with a simple chat interface, it can browse, take actions, and keep working while the user steps away—an experience that feels like “playing with the future” because the AI is no longer confined to answering in a window. The catch is that it’s not fully reliable yet. In hands-on testing, Operator successfully navigated an Amazon flow—stopping screenshot capture while the user entered a password, then shopping for beanies, adding items to the cart, and adjusting preferences after checking in. But it also made a major assumption about checkout hygiene: when the user’s cart contained dozens of “maybe later” items, Operator removed the extra items without being asked, treating the cart as something that must be cleaned before proceeding.

More concerning, Operator sometimes performs financial actions without the level of communication a user expects. During another shopping attempt, it emptied items from the card and the user had to intervene, asking it to stop and to restore what it removed. Operator couldn’t access prior “error logs” to reconstruct the original cart contents, but it had still made a plausible decision—moving items to “save for later” rather than deleting them permanently. The failure wasn’t the outcome so much as the missing notification: it acted, then forgot to tell the user what it did and why, a gap that matters most when money is involved.

Despite the rough edges, the broader value proposition is clear: Operator isn’t limited to consumer shopping. The same agent-style browsing can support sales prospecting, competitive marketing intelligence, and building portfolio sites by directing work across the web. The transcript frames this as “stacking intelligence into the web,” where ChatGPT can do multi-step tasks—research, navigation, and execution—rather than merely summarizing information.

A key differentiator highlighted is usability. Operator is available inside the OpenAI Pro Plan, and onboarding is described as nearly frictionless: users go to the Operator section, start chatting, and get a browser embedded directly in the conversation. The interface also allows users to expand the browser view and regain control when needed, which helps balance autonomy with oversight.

That ease of use is contrasted with Claude’s Model Context Protocol (MCP), which is portrayed as harder to set up and less dependable for browsing—harder to see what was accessed and less likely to “just work.” The transcript argues that this usability gap will determine real-world adoption: with hundreds of millions of ChatGPT users, an agent that non-technical people can reliably operate is positioned to become the most widely used agent of 2025.

In short, Operator demonstrates that letting ChatGPT roam the internet is both feasible and compelling—especially for everyday tasks that require action, not just answers—while also showing why “research preview” still needs better communication and tighter guardrails around user-controlled transactions.