MacAgentBench - Development: leaderboard
The 44 development tasks in VS Code and Tmux of MacAgentBench (Fu et al., 2026), run in a containerized macOS environment and scored by deterministic rule-based checks of the final state. Pass@1 in percent with the benchmark baseline pure-GUI agent (screenshots and mouse and keyboard actions); higher is better.
Metric: Pass@1 (%; share of the 44 Development tasks solved on a single attempt, deterministic rule-based checks; the benchmark's baseline pure-GUI agent (screenshots and mouse/keyboard actions only), at most 50 steps). Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 77.3 | #121 |
| 2 | Claude Opus 4.6 | 65.9 | #100 |
| 3 | Gemini 3.1 Pro (Preview) | 59.1 | #91 |
| 4 | Qwen 3 VL 32B (Thinking) | 9.1 | #557 |
| 5 | Qwen 3 VL 235B A22B | 6.8 | #477 |
| 6 | Qwen 3 VL 8B (Thinking) | 2.3 | #887 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=macagentbench-development · How It Works · Data refreshed daily, snapshot 2026-11-12.