# MacAgentBench Leaderboard

> Last updated AUG 31, 2026

Tracking reported computer-use agent performance on MacAgentBench, combining the official benchmark leaderboard with additional results found by Yutori Scout.

Official results come from the MacAgentBench leaderboard. The official leaderboard has not been updated since Jun 21, 2026. We’ve set up a [Scout](<https://yutori.com/scouting>) to monitor for additional results and catch results it has missed.

## Reported results

| Model | Organization | Agent | Score |
| --- | --- | --- | --- |
| [Navigator n2](<https://yutori.com/blog/introducing-n2>) | Yutori | – | 83.1%<sup>1</sup> |
| Claude Opus 4.6 | Anthropic | OpenClaw | 73.7% |
| Claude Opus 4.6 | Anthropic | Agent-S3 | 66.9% |
| GPT-5.5 | OpenAI | – | 66.7%<sup>2</sup> |
| Gemini 3.1 Pro | Google | OpenClaw | 63.3% |
| GPT-5.4 | OpenAI | OpenClaw | 60.7% |
| GPT-5.4 | OpenAI | Agent-S3 | 58.9% |
| GPT-5.4 | OpenAI | Baseline | 58.4% |
| Claude Opus 4.8 | Anthropic | – | 58.4%<sup>2</sup> |
| Gemini 3.1 Pro | Google | Agent-S3 | 54.3% |
| Qwen3VL-235B-A22B | Alibaba | Agent-S3 | 46.7% |
| Qwen3VL-235B-A22B | Alibaba | OpenClaw | 46.0% |
| Claude Opus 4.6 | Anthropic | Baseline | 39.2% |
| Qwen3VL-32B | Alibaba | Agent-S3 | 37.4% |
| Gemini 3.1 Pro | Google | Baseline | 34.2% |
| Qwen3VL-32B | Alibaba | OpenClaw | 33.7% |
| Qwen3VL-8B | Alibaba | Agent-S3 | 28.6% |
| Qwen3VL-8B | Alibaba | OpenClaw | 26.9% |
| Qwen3VL-235B-A22B | Alibaba | Baseline | 21.6% |
| Qwen3VL-32B | Alibaba | Baseline | 21.3% |
| OpenCUA-32B | XLang | Baseline | 18.8% |
| GUI-Owl-1.5-32B | Alibaba | Baseline | 17.2% |
| OpenCUA-7B | XLang | Baseline | 15.1% |
| Qwen3VL-8B | Alibaba | Baseline | 14.5% |
| InternVL3.5-14B | OpenGVLab | Agent-S3 | 13.5% |
| UI-TARS-72B-DPO | ByteDance | Baseline | 13.2% |
| ScaleCUA-32B | OpenGVLab | Baseline | 10.5% |
| GUI-Owl-1.5-8B | Alibaba | Baseline | 10.2% |
| UI-TARS-1.5-7B | ByteDance | Baseline | 9.8% |
| InternVL3.5-8B | OpenGVLab | Agent-S3 | 9.6% |
| ScaleCUA-7B | OpenGVLab | Baseline | 6.7% |
| InternVL3.5-14B | OpenGVLab | Baseline | 6.4% |
| InternVL3.5-8B | OpenGVLab | Baseline | 4.7% |

<sup>1</sup> Navigator n2’s MacAgentBench score is reported in the [launch post](<https://yutori.com/blog/introducing-n2>).

<sup>2</sup> [Claude Opus 4.8 and GPT-5.5 results reported by Alibaba in the Qwen-CUA technical report.](<https://github.com/xlang-ai/Qwen-CUA/blob/main/paper/Qwen-CUA.pdf>)

## About the benchmark

MacAgentBench is a macOS desktop-agent benchmark with 676 tasks across 25 applications. It uses deterministic rule-based evaluation, fine-grained checkpoints, and model comparisons across Baseline, Agent-S3, and OpenClaw frameworks.

## Sources

- [Paper (arXiv 2606.22557)](<https://arxiv.org/abs/2606.22557>)
- [Code (GitHub · JetAstra)](<https://github.com/JetAstra/MacAgentBench>)
- [Environment (Hugging Face)](<https://huggingface.co/JetLM/OpenClaw-macOS>)
- [Official leaderboard](<https://jetastra.github.io/MacAgentBench/>)
- [Navigator n2 launch post](<https://yutori.com/blog/introducing-n2>)

Official source revision (SHA-256): `20d9e2f8ede68266a3e60f2f20c4ff5644a4bb06bd9a46af8688413c7b3ca9e0`