InfraBench - Best-of-3 Perfect Tasks: leaderboard
InfraBench (arXiv 2608.11234v2, Table 3): share of the 12 infrastructure tasks on which at least one of three attempts by an agent CLI plus pinned model configuration earns a perfect difficulty-weighted verifier score. Percent in steps of about 8.3 points; higher is better.
Metric: Tasks with a perfect best-of-three attempt (%; of 12). Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 | 83.3 | #20 |
| 2 | Grok 4.5 | 83.3 | #54 |
| 3 | Claude Opus 4.8 | 75 | #39 |
| 4 | Claude Opus 5 | 75 | #16 |
| 5 | Claude Sonnet 5 | 75 | #65 |
| 6 | Gemini 3.6 Flash | 75 | #57 |
| 7 | DeepSeek V4 Flash | 66.7 | #155 |
| 8 | Gemini 3.5 Flash | 66.7 | #66 |
| 9 | Composer 2.5 | 66.7 | #130 |
| 10 | Gemini 3.1 Pro (Preview) | 58.3 | #64 |
| 11 | Claude Sonnet 4.6 | 50 | #94 |
| 12 | DeepSeek V4 Pro | 50 | #112 |
| 13 | MiMo-V2.5 | 50 | #196 |
| 14 | Gemini 3.1 Flash Lite | 41.7 | #200 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=infrabench-best-of-3-perfect-tasks · How It Works · Data refreshed daily, snapshot 2026-11-05.