EC2-Bench - Completion: leaderboard
EC2-Bench (EC²-Bench, OpenBMB; arXiv 2610.17816): 128 executable digital-workspace tasks (office work, search, safety, devops, data analysis, automation) in one OpenClaw agent harness with mock services, adapted from Claw-Eval, WildClawBench and other agent suites. Only single models running alone are listed; the paper's edge-cloud collaboration systems are not. Score: weighted completion score (0-100) from rule, environment-state and rubric checks with a GPT-5.4-mini judge, mean of three runs. Higher is better. Source: https://arxiv.org/abs/2610.17816
Metric: Task completion score (0-100; weighted rule, environment-state and rubric LLM-judge checks per task, GPT-5.4-mini judge, mean of 3 runs over 128 OpenClaw workspace tasks). Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.3 | 82.7 | #39 |
| 2 | DeepSeek V4.1 Flash | 81.42 | #35 |
| 3 | DeepSeek V4 Pro | 72.9 | #100 |
| 4 | Qwen 3.7 Plus | 71.34 | #91 |
| 5 | Qwen 3.5 27B | 61.79 | #185 |
| 6 | Qwen 3.5 9B | 39.12 | #377 |
| 7 | Qwen 3.5 2B | 12.75 | #1145 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ec2-bench-completion · How It Works · Data refreshed daily, snapshot 2026-10-19.