Benchmarks
326 sources · 1,256 models × 2,956 benchmarks · 154,484 scoresCua's Cua Driver Bench v2 on macOS 26 while a simulated user keeps typing in a window in front: eight of its ten desktop tasks (the two hover tasks run only in the foreground), done only through a computer-use driver's tools. A trial passes when the task's checker confirms the apps' final state, no route out of the GUI was used and none of eight disturbances reached the user (lost focus, a window raised over theirs, lost typing, pointer, clipboard, Space, a dialog or deactivation); each task runs five seeded trials and the rate is passes over trials.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | — · Cua Driver Bench on AgentEnv; Claude Code 2.1.280 with no --effort passed and its built-in tools limited to Skill and Read; trycua's cua-driver MCP tools alone; 45 turns and 360 seconds per trial | 57.5 | |
| 2 | Claude Sonnet 5.5 | — · Cua Driver Bench on AgentEnv; Claude Code 2.1.280 with no --effort passed and its built-in tools limited to Skill and Read; trycua's cua-driver MCP tools alone; 45 turns and 360 seconds per trial | 57.5 | |
| 3 | Claude Haiku 5.5 | — · Cua Driver Bench on AgentEnv; Claude Code 2.1.280 with no --effort passed and its built-in tools limited to Skill and Read; trycua's cua-driver MCP tools alone; 45 turns and 360 seconds per trial | 45.0 | |
| 4 | GPT-6.1 Sol | medium · Cua Driver Bench on AgentEnv; Codex CLI 0.147.0 with every built-in Codex tool off; trycua's cua-driver MCP tools alone, whose screenshots Codex doesn't pass to the model; stopped after 45 tool calls or 360 seconds per trial | 40.0 |