EC2-Bench - Pass^3: leaderboard
EC2-Bench (arXiv 2610.17816): 128 OpenClaw workspace tasks run three times per model, single models only. Score: Pass^3, the share of tasks (%) whose completion score exceeds 0.7 in all three runs, a measure of reliable success. Higher is better. Source: https://arxiv.org/abs/2610.17816
Metric: Pass^3 (%; share of the 128 tasks whose completion score exceeds 0.7 in all three runs). Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.3 | 71.88 | #39 |
| 2 | DeepSeek V4.1 Flash | 71.88 | #35 |
| 3 | DeepSeek V4 Pro | 53.12 | #100 |
| 4 | Qwen 3.7 Plus | 53.12 | #91 |
| 5 | Qwen 3.5 27B | 30.47 | #185 |
| 6 | Qwen 3.5 9B | 9.38 | #377 |
| 7 | Qwen 3.5 2B | 2.34 | #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-pass-pow-3 · How It Works · Data refreshed daily, snapshot 2026-10-19.