On Thursday, OpenAI unveiled a research preview of "Operator, " a web automation tool powered by a new AI model called Computer-Using Agent (CUA). This tool enables users to control computers through a visual interface, performing tasks by interacting with on-screen elements like buttons and text fields, similar to human behavior. Currently, Operator is accessible to subscribers of the $200 monthly ChatGPT Pro plan at operator. chatgpt. com, with plans to extend access to Plus, Team, and Enterprise users in the future. OpenAI aims to integrate these features into ChatGPT and provide the CUA through an API for developers. Operator monitors screen content in real-time, executing commands through simulated keyboard and mouse actions. It analyzes screenshots to understand the computer's state, which allows it to make decisions on clicking, typing, and scrolling accordingly. This release aligns with a trend among tech companies exploring "agentic" AI systems capable of acting on user behalf. Google introduced Project Mariner in December 2024 for automated web tasks, while Anthropic launched a tool called "Computer Use" in October 2024 for similar purposes. Simon Willison, an AI researcher, noted that the Operator interface resembles Anthropic’s Claude Computer Use demo, featuring a chat panel alongside an interactive visual interface. The CUA operates by capturing screenshots, analyzing them to assess pixel data with GPT-4o's vision capabilities, deciding on necessary actions, and executing virtual inputs. This feedback loop allows it to rectify errors and manage complex tasks across various applications.

During use, a miniature browser window displays Operator's actions. Despite its capabilities, the technology is still evolving and has limitations, excelling in repetitive web tasks but struggling with unfamiliar interfaces and complex text editing—reporting a mere 40 percent success rate in internal tests. Operator achieved an 87 percent success rate on the WebVoyager benchmark but dropped to 58. 1 percent on the WebArena benchmark for offline training sites. Its performance on the OSWorld benchmark was 38. 1 percent. OpenAI is seeking user feedback to enhance the system's functionality, understanding that CUA won’t be reliable in all scenarios and plans to improve its versatility over time. Privacy and safety are key concerns, given that Operator can observe and control user activities. OpenAI has implemented safety measures requiring user confirmation for sensitive actions, with browsing limits established to prevent access to certain site categories such as gambling and adult content. OpenAI also integrates real-time moderation to prevent subversion attempts, though Willison expressed skepticism about its long-term security due to the potential for new threats. Despite proactive measures, OpenAI admits in its documentation that challenges remain due to modeling complexities and evolving adversarial threats. Privacy concerns arise from the fact that Operator sends screenshots to OpenAI's cloud servers; users are encouraged to trust the company's privacy measures, which allow data opt-out for model training, one-click data deletion, and session management for sensitive tasks. Willison advised users to initiate fresh sessions for separate tasks to protect their credentials and urged caution when providing payment details, recommending a quick cleanup following such actions.