UniClawBench: leaderboard
UniClawBench (HKU MMLab, 2026) runs proactive agents on 400 bilingual real-world tasks in live Docker containers under the OpenClaw framework, with up to two simulated user follow-ups and a hidden GPT-5.4 supervisor that grades step checkpoints. Score is the pass rate over all tasks, the mean of the English and Chinese subsets, in percent; higher is better.
Metric: Pass rate (%; tasks reaching a final pass state, mean of the English and Chinese subsets) on 400 bilingual real-world tasks run live in Docker containers under the OpenClaw framework (v2026.3.11), with up to two simulated user follow-ups and a hidden GPT-5.4 Codex supervisor that grades step checkpoints. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 47.5 | #56 |
| 2 | Claude Sonnet 4.6 | 45.5 | #161 |
| 3 | GPT-5.4 | 40.7 | #123 |
| 4 | Kimi K2.6 | 36.2 | #178 |
| 5 | Gemini 3.1 Pro (Preview) | 32.5 | #94 |
| 6 | Qwen 3.5 Plus | 31.8 | #232 |
| 7 | Gemini 3 Flash | 28.2 | #164 |
| 8 | GPT-5.4 Mini | 21.5 | #389 |
| 9 | Gemini 3.1 Flash Lite | 15.5 | #364 |
| 10 | GPT-4.1 | 15.2 | #495 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=uniclawbench · How It Works · Data refreshed daily, snapshot 2026-11-13.