UniClawBench - Average Score: leaderboard

Metric: Average checkpoint score (%; mean completed share of the task checkpoints, English and Chinese subsets averaged) 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 December 2026. 10 models tracked.

Top models

#ModelScore
1GPT-5.477.4
2Claude Sonnet 4.676.3
3Qwen 3.5 Plus73.1
4Gemini 3.1 Pro (Preview)72.7
5Kimi K2.670.9
6Claude Opus 4.870.2
7Gemini 3 Flash69.8
8GPT-5.4 Mini61.6
9Gemini 3.1 Flash Lite54.3
10GPT-4.148.8

Interactive version: theaggregate.ai/benchmark?slug=uniclawbench-average-score · How It Works · Data refreshed daily, snapshot 2026-09-29.