Computer-Using Agent (CUA)
2025-01-23 · Publisher : OpenAI · Domain : Artificial Intelligence
OpenAI introduces Computer-Using Agent (CUA), a model that combines GPT-4o's vision with reinforcement-learning-trained reasoning to operate a graphical interface. It powers the Operator agent.
What changes
- CUA reads screenshots, reasons through chain of thought, then acts with a virtual mouse and keyboard, without going through specialized APIs.
- Three-step loop: perception, reasoning, action, with a request for user confirmation for sensitive actions (credentials, CAPTCHA).
- First OpenAI model dedicated to computer use, delivered as a research preview in Operator.
- API access will come on March 11, 2025 with the computer use tool (computer-use-preview model, see the dedicated entry).
### Specifications - Inputs: screenshots (pixels); outputs: click, scroll and typing actions. - Availability at announcement: Operator (operator.chatgpt.com), ChatGPT Pro subscribers in the United States, as a research preview. - Context window, maximum output and knowledge cutoff: not disclosed in the announcement. ### Pricing Pricing not disclosed at launch (access through the Pro subscription). ### Benchmarks Scores published by the vendor. | Benchmark | CUA | Previous state of the art | Human | | --- | --- | --- | --- | | OSWorld (full computer use) | 38.1% | 22.0% | 72.4% | | WebArena (web navigation) | 58.1% | 36.2% (universal interface), 57.1% (specialized web agents) | 78.2% | | WebVoyager (online websites) | 87.0% | 56.0% (universal interface), 87.0% (specialized web agents) | not disclosed | ### Key takeaways CUA lays the groundwork for automating tasks through graphical interfaces, useful where no API exists. The success rate remains far from human level on complex tasks, which calls for supervision.
Links
Tags : LLM, OpenAI, Agent, Computer Use, Operator