OpenAI launches Operator
Built on a new Computer-Using Agent model layered on GPT-4o, it scored 38.1% on OSWorld against a 72.4% human baseline, and launched to $200-a-month Pro subscribers only.
- Models & capabilities
- Notable
OpenAI released Operator, a research-preview agent that controlled a dedicated web browser to complete tasks such as booking travel, ordering groceries and making restaurant reservations from text instructions. It launched to ChatGPT Pro subscribers — OpenAI’s $200-a-month tier — in the United States only, with wider rollout to Plus, Team and Enterprise plans, and to Europe, promised for later.
Operator ran on a new model OpenAI called Computer-Using Agent (CUA), built on GPT-4o’s vision capabilities and combined with reasoning techniques similar to o1 and o3. Rather than working through an API, CUA looked at screenshots of a webpage, identified the available actions, and interacted the way a person would — typing, clicking, scrolling — which let it work with sites that had no integration with OpenAI at all, though the company also announced launch partnerships with services including DoorDash, Instacart, eBay and Priceline.
OpenAI reported CUA scored 38.1% on OSWorld, a benchmark of general computer-use tasks, against 22.0% for Anthropic’s Computer Use and a human baseline of 72.4%; on WebVoyager, a browser-task benchmark, it scored 87%, ahead of Google DeepMind’s Mariner at 83.5%. The company was explicit the system remained unreliable: it could get stuck on CAPTCHAs and complex interfaces, refused to send emails or delete calendar events, and required user confirmation before any action with real-world consequences such as a purchase.
Operator launched alongside its system card, which detailed the new risks — chiefly prompt injection from web content — that an agent able to act on the open web introduced. OpenAI later folded Operator’s capabilities into a broader agent product and shut the standalone service down.