Dynamic Insight Beating AI News: Cua, in collaboration with Scale AI, has launched the Cua Driver Bench for computer operation evaluation. It requires AI Agents to operate software in the background without seizing the user's mouse or keyboard, nor popping up windows to interrupt work.
In macOS testing, Cua Driver achieved a full success rate of 58%, Codex Computer Use reached 33%, Vercel agent-desktop scored 20%, and browser-use only managed 3%.
Cua's advantage stems from direct manipulation of background windows. Conventional computer operation tools typically simulate human clicks and typing, requiring the target window to be brought to the foreground. Cua, however, leverages macOS's accessibility APIs to directly locate buttons and input fields within background applications and execute actions. It can also capture individual window screenshots without needing to bring the window to the top.
The disparity is especially evident in browser-use. In 40 tests, it successfully completed 32 tasks, but disrupted the user in 31 instances—resulting in only one fully successful test. Although Cua completed only 25 tasks, 23 of them did not disturb the user, and it never seized keyboard focus during testing.
All tools used Claude Sonnet 5.5; the comparison focuses on computer operation tools rather than the models themselves. Tests covered tasks involving browsers, email, calendars, spreadsheets, and more, each requiring completion while the user continued typing.