MacAgentBench (Agent-S3): leaderboard
macOS desktop automation on MacAgentBench (Fu et al., 2026): 676 tasks in 25 applications, most mixing GUI and command-line work, run in a containerized macOS Tahoe environment and scored by deterministic rule-based checks of the final system state. Pass@1 in percent with the Agent S3 multi-agent framework, at most 50 steps per task; higher is better.
Metric: Pass@1 (%; share of the 676 macOS desktop tasks in 25 applications solved on a single attempt, judged by deterministic rule-based state checks; the Agent S3 multi-agent framework, at most 50 steps). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 66.9 | #100 |
| 2 | GPT-5.4 | 58.9 | #121 |
| 3 | Gemini 3.1 Pro (Preview) | 54.3 | #91 |
| 4 | Qwen 3 VL 235B A22B | 46.7 | #477 |
| 5 | Qwen 3 VL 32B (Thinking) | 37.4 | #557 |
| 6 | Qwen 3 VL 8B (Thinking) | 28.6 | #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-agent-s3 · How It Works · Data refreshed daily, snapshot 2026-11-12.