Cua conducts a practical test of AI operating computers via backend: its own pass rate stands at 58%, while browser-use only has a 3% pass rate.

Beating AI News Flash: Cua has partnered with Scale AI to launch Cua Driver Bench, a computer operation benchmark. The benchmark requires AI agents to operate software in the background without seizing the user’s mouse, keyboard, or popping up windows to disrupt work. In macOS tests, Cua Driver achieved a full pass rate of 58%, compared to 33% for Codex Computer Use, 20% for Vercel agent-desktop, and just 3% for browser-use. Cua’s advantage stems from its ability to directly operate background windows. Conventional computer operation tools typically simulate human clicks and typing, requiring the target window to be brought to the foreground. Cua, by contrast, uses macOS’s accessibility APIs to directly locate buttons and input fields in background apps and execute operations. It can also capture individual window screenshots without moving the window to the top layer. The gap is especially notable for browser-use: in 40 tests, it successfully completed 32 tasks, but 31 of them interrupted the user, leading to only 1 fully passed test. Cua, meanwhile, completed 25 tasks, with 23 cases of no user disruption and zero instances of seizing keyboard focus during testing. All tools used Claude Sonnet 5.5; the comparison evaluated computer operation tools, not the underlying models. Tests covered tasks including browsers, email, calendars, and spreadsheets, each requiring completion while the user was actively typing.