OpenAI's Operator can browse the web for you — what it gets right

OpenAI just let its AI leave the chat window and start clicking around the actual web. The company launched Operator, a research preview of an agent that uses its own browser to fill out forms, order groceries, and even create memes — no custom APIs required. For now, it’s only available to Pro users in the U.S., with Plus, Team, and Enterprise access promised down the line. The whole thing runs on a new model called Computer-Using Agent, or CUA, which combines GPT‑4o’s vision smarts with reinforcement learning to “see” screenshots and “interact” with buttons, menus, and text fields the way a human would — typing, clicking, scrolling. OpenAI says CUA already notches state-of-the-art results on the WebArena and WebVoyager benchmarks, two standard tests for browser-based tasks.

The pitch is straightforward: give it a repetitive online chore and it’ll execute. But the safety rails are telling. Operator is trained to hand control back to users for anything involving logins, payment details, or CAPTCHAs — the messy parts of the web where an autonomous agent could do real damage. Users can also take over the remote browser at any point, and they can set custom instructions for specific sites, like airline preferences on Booking.com. The system can even juggle multiple tasks at once by spinning up new conversations, so you could theoretically order a custom mug on Etsy while it books a campsite on Hipcamp.

OpenAI is framing this as a collaborative rollout, not a finished product. The company name-dropped partnerships with DoorDash, Instacart, OpenTable, Priceline, StubHub, Thumbtack, Uber, and others — a signal that it wants buy-in from the platforms whose interfaces Operator will be scraping. There’s also a public-sector angle: the City of Stockton is working with OpenAI to explore how the agent might streamline enrollment in city services. Stockton’s CIO Harry Black called Operator “a technological breakthrough that makes processes like ordering groceries incredibly easy,” though it’s worth noting that’s a quote from a press release, not an independent review.

I’ll say what a lot of people in the research community are thinking: browser agents are having a moment, but the gap between a benchmark score and actual reliability is a chasm. WebArena results are nice, but real websites change their layouts constantly and break brittle automations. The decision to start with a limited research preview — and to sunset the standalone Operator site in favor of baking everything into ChatGPT’s “agent mode” — suggests OpenAI knows this technology needs a lot of real-world mileage before it’s ready for prime time. The question isn’t whether an AI can click a button. It’s whether it can recover gracefully when the button moves.