Source: OpenAI, Computer-Using Agent (CUA)
OpenAI's January 2025 announcement of CUA, the GPT-4o-based RL-trained model behind Operator, with OSWorld, WebArena and WebVoyager results.
Contents
- What: vendor research announcement
- Author/publisher: OpenAI
- Date: 2025-01-23
- Original: Computer-Using Agent
Summary
CUA combines GPT-4o vision with reasoning trained by reinforcement learning to operate GUIs from pixels, without OS- or site-specific APIs. It powered the Operator research preview.
Key claims
- OSWorld 38.1% (prior SOTA 22.0%, human 72.4%); WebArena 58.1% (human 78.2%); WebVoyager 87%.
- WebVoyager tasks are relatively simple; a large gap to humans remained on complex WebArena tasks.
- Launched in Operator for US Pro users; API access planned.
- OpenAI did not expect reliable performance in all scenarios yet.
Note: the page returned 403 to direct fetch; claims taken from search-result extracts of the page and launch coverage.