LiveClawBench: leaderboard

Agent performance on LiveClawBench (2026), personal-assistant agent tasks in reproducible full-stack mock services (stateful web applications, browser, APIs and shell) with implicit goals and runtime perturbations, three runs per case scored by case verifiers and a rubric judge. Average reward over all 134 cases, from 0 to 100; higher is better.

Metric: Average reward (0-100) over three runs on all 134 cases, 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 February 2027. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Kimi K2.7 Code76#203
2GLM-5.174.7#198
3GPT-5.574.5#54
4GLM-5.272.9#150
5MiniMax-M371.4#280
6Qwen 3.6 Plus70.6#238
7MiMo-V2.5-Pro70.6#284
8Qwen 3.6 Flash70.2#448
9Claude Opus 4.868.8#65
10Kimi K2.668.6#192
11MiniMax-M2.768.4#539
12DeepSeek V4 Pro65.6#196
13DeepSeek V4 Flash60.5#267
14Qwen 3.5 Plus57#242
15Qwen 3.6 27B54.5#351

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

Example tasks

I just had a chicken salad and a glass of milk for lunch. Please log this into my diet logging app at http://localhost:5003/ — open it in your browser and add these to my lunch for today.

Source

I recently booked a flight and received a booking notification in my company email system (http://localhost:5174/, open it in browser). Please help me complete the check-in and seat selection. I'd like a window seat.

Source

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