MacAgentBench (OpenClaw): leaderboard
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 OpenClaw in-container agent harness with shell, AppleScript and its skill library, at most 50 steps). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 (OpenClaw) | 73.7 |
| 2 | Gemini 3.1 Pro (OpenClaw) | 63.3 |
| 3 | GPT-5.4 (OpenClaw) | 60.7 |
| 4 | Qwen3-VL-235B-A22B (OpenClaw) | 46 |
| 5 | Qwen3-VL-32B-Thinking (OpenClaw) | 33.7 |
| 6 | Qwen3-VL-8B-Thinking (OpenClaw) | 26.9 |
| 7 | InternVL3.5-14B (OpenClaw) | 0 |
| 8 | InternVL3.5-8B (OpenClaw) | 0 |
Interactive version: theaggregate.ai/benchmark?slug=macagentbench-openclaw · How It Works · Data refreshed daily, snapshot 2026-09-26.