Cua tested AI backend computer operation: its own success rate 58%, browser-use only 3%

Beating AI News Flash: Cua has partnered with Scale AI to launch Cua Driver Bench, a computer operation benchmark. It requires AI Agents to operate software in the background without seizing the user's mouse or keyboard, and without popping up windows that interrupt work.

In macOS testing, Cua Driver achieved a full pass rate of 58%, Codex Computer Use reached 33%, Vercel agent-desktop reached 20%, and browser-use only 3%.

Cua's advantage comes from directly operating background windows. Ordinary computer operation tools typically simulate human clicks and typing, requiring the target window to be brought to the foreground. Cua instead uses macOS accessibility interfaces to directly locate buttons and input fields in background applications and perform operations. It can also capture window images individually without bringing the window to the top layer.

The gap is especially pronounced with browser-use. In 40 tests, it successfully completed 32 tasks, but 31 of those interfered with the user, leaving only 1 full pass. Although Cua completed only 25, 23 of those did not disturb the user, and it never seized keyboard focus during testing.

All tools used Claude Sonnet 5.5; the comparison is of computer operation tools, not the models themselves. Tests covered tasks such as browsers, email, calendars, and spreadsheets, each requiring completion while the user continuously typed.