CuaDriverBench AgentEnv - macOS / agent-desktop 1.2.0 / Foreground: leaderboard
CuaDriverBench on AgentEnv tests GUI-only computer use while a human shares the desktop. The foreground condition gives the agent the desktop alone; background requires completing the task without disturbing the user’s focus, windows, pointer, clipboard or typing. A full pass requires task completion and GUI-only behavior, plus leaving the user undisturbed in background. This reader retains all 21 cohorts in the publisher’s current board snapshot: 5,370 trials, grouped separately by operating system, GUI driver, driver version and condition. Five trials per available task are retained, including failures. Native model API identity, CLI harness/version and reported reasoning effort are bound to the same immutable trial dataset. Each model keeps its native available-task denominator; Omarchy uses its own BenchLab subset. Score is 100 × full passes / trials from the native parquet, rather than the rounded README display, task-completion-only rate, duration or cost. Historical macOS 0.38.0 uses a separate reader. No Haiku/Fable or model-generation identities are merged. This board uses macOS, agent-desktop 1.2.0, foreground condition; higher is better.
Metric: Full GUI-only pass rate (%). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 2 models tracked. Updated 14 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5.5 (Harness: claude-code v2.1.280; agent-desktop v1.2.0; macOS) | 70 | |
| 2 | Claude Haiku 5.5 (Harness: claude-code v2.1.280; agent-desktop v1.2.0; macOS) | 64 |
Interactive version: theaggregate.ai/benchmark?slug=cuadriverbench-agentenv-macos-agent-desktop-1-2-0-foreground · How It Works · Data refreshed daily, snapshot 2026-11-16.