ScrambleToolBench - Execution Windows - Tasks Solved: leaderboard
ScrambleToolBench (arXiv 2608.02358v1, Table 2): agents work through five sequential tasks using tools whose names and argument schemas are obfuscated, so tool behaviour must be discovered by probing. Mean tasks solved per episode (0-5, 20 episodes) when some action sequences must finish within a strict step limit. Higher is better.
Metric: Average tasks solved per five-task episode (0-5). Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 4.8 | #64 |
| 2 | Claude Sonnet 5 | 4.7 | #65 |
| 3 | Gemini 3.5 Flash | 4.65 | #66 |
| 4 | Claude Sonnet 4.5 | 4 | #148 |
| 5 | Qwen 3.6 27B | 2.9 | #182 |
| 6 | Gemini 2.5 Pro | 2.8 | #159 |
| 7 | Gemini 2.5 Flash | 1.45 | #255 |
| 8 | GPT-5.4 | 1.25 | #83 |
| 9 | Gemma 4 31B | 0.75 | #212 |
| 10 | Qwen 3.6 35B A3B | 0.3 | #238 |
| 11 | Gemma 4 26B A4B | 0.3 | #360 |
| 12 | GPT-5.4 Mini | 0.1 | #228 |
| 13 | Gemini 3.1 Flash Lite | 0.1 | #200 |
| 14 | GPT-OSS-120B | 0 | #346 |
| 15 | GPT-OSS-20B | 0 | #523 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=scrambletoolbench-execution-windows-tasks-solved · How It Works · Data refreshed daily, snapshot 2026-11-05.