RealClawBench - Pass@3: leaderboard
RealClawBench (2026) turns real OpenClaw developer-agent sessions into 281 reproducible tasks with reconstructed workspaces, run in the shared OpenClaw agent runtime and scored by deterministic task-specific verifiers over three runs. This board is pass@3, the share (%) of tasks solved in at least one of three runs; higher is better.
Metric: Pass@3 (%): share of the 281 tasks solved in at least one of three runs, on 281 released tasks reconstructed from real OpenClaw developer-agent sessions, run in the shared OpenClaw agent runtime (file read, write, edit, search and shell tools, 600-second timeout, temperature 1.0) and scored by case-specific deterministic Python verifiers; mean of three independent runs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 76.2 | #188 |
| 2 | GPT-5.5 | 75.1 | #49 |
| 3 | Claude Opus 4.7 | 75.1 | #79 |
| 4 | MiMo-V2.5-Pro | 74.7 | #279 |
| 5 | Kimi K2.6 | 73 | #178 |
| 6 | DeepSeek V4 Flash | 71.2 | #258 |
| 7 | GLM-5.1 | 70.8 | #179 |
| 8 | Claude Sonnet 4.6 | 70.5 | #161 |
| 9 | Gemini 3.1 Pro (Preview) | 69 | #94 |
| 10 | MiniMax-M2.7 | 68.3 | #518 |
| 11 | Qwen 3.6 Plus | 67.6 | #222 |
| 12 | Claude Opus 4.6 | 67.3 | #102 |
| 13 | GPT-OSS-120B | 57.7 | #700 |
| 14 | Gemma 4 31B | 51.6 | #394 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=realclawbench-pass-3 · How It Works · Data refreshed daily, snapshot 2026-11-13.