InfraBench - Attempt Pass@1: leaderboard
InfraBench (arXiv 2608.11234v2, Table 3): share of the 36 attempts (12 infrastructure tasks x 3) in which an agent CLI plus pinned model configuration earns a perfect difficulty-weighted verifier score, at vendor-default CLI settings. Percent; higher is better. An attempt rate, not a sampled pass@k estimate.
Metric: Attempts with a perfect score (%). Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 | 77.8 | #20 |
| 2 | Grok 4.5 | 75 | #54 |
| 3 | Claude Opus 5 | 72.2 | #16 |
| 4 | Gemini 3.6 Flash | 66.7 | #57 |
| 5 | Claude Opus 4.8 | 63.9 | #39 |
| 6 | Gemini 3.5 Flash | 63.9 | #66 |
| 7 | Claude Sonnet 5 | 63.9 | #65 |
| 8 | Composer 2.5 | 58.3 | #130 |
| 9 | Gemini 3.1 Pro (Preview) | 52.8 | #64 |
| 10 | DeepSeek V4 Flash | 47.2 | #155 |
| 11 | MiMo-V2.5 | 38.9 | #196 |
| 12 | Claude Sonnet 4.6 | 36.1 | #94 |
| 13 | DeepSeek V4 Pro | 36.1 | #112 |
| 14 | Gemini 3.1 Flash Lite | 33.3 | #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-attempt-pass-1 · How It Works · Data refreshed daily, snapshot 2026-11-05.