LiveClawBench - Hard: leaderboard

Agent performance on the 19 Hard cases of LiveClawBench (2026), personal-assistant agent tasks in reproducible full-stack mock services; a case is Hard when the anchor model DeepSeek-V4-Pro earns a mean reward of 0.3 or less on it. Average reward over three runs, from 0 to 100; higher is better.

Metric: Average reward (0-100) over three runs on the 19 Hard cases (DeepSeek-V4-Pro anchor reward 0.3 or less), OpenClaw-style personal-assistant tasks run in reproducible full-stack mock services (22 stateful web applications, browser, APIs and shell), three runs per case, rewards from case verifiers and a DeepSeek-V3.2 rubric judge for open-ended cases; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GLM-5.127.7#198
2Kimi K2.7 Code25.9#203
3MiniMax-M325.7#280
4GLM-5.224.3#150
5GPT-5.522#54
6MiniMax-M2.719.5#539
7Kimi K2.618.3#192
8MiMo-V2.5-Pro17.2#284
9Qwen 3.6 Flash14.3#448
10Qwen 3.6 Plus13.2#238
11DeepSeek V4 Flash13#267
12Qwen 3.5 Plus12.8#242
13Claude Opus 4.810.2#65
14Qwen 3.5 Flash5#441
15Qwen 3.6 27B4.7#351

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=liveclawbench-hard · How It Works · Data refreshed daily, snapshot 2026-11-21.