Claude Sonnet 5.5 Shows Mixed Desktop Control

The results favor a hybrid computer-use architecture rather than a screenshot-versus-tree winner.

• –Accessibility trees provide faster, more precise control of standard UI elements.
• –Screenshot input remains valuable for canvas apps, custom controls, and visual state verification.
• –The 83% versus 75% gap is directionally interesting, but the sample is too small for broad reliability claims.
• –Developers should route structured widgets through accessibility APIs while retaining screenshot fallback, consistent with Anthropic’s computer-use guidance ([official docs](https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool)).
• –Broader testing across apps, resolutions, permissions, and long-horizon tasks will determine whether this advantage generalizes.