InfraBench - Attempt Pass@0.5: 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 reaches at least half of the difficulty-weighted verifier score, i.e. substantially solves the task. Percent; higher is better.
Metric: Attempts scoring at least 0.5 (%). Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 | 91.7 | #20 |
| 2 | Claude Opus 4.8 | 86.1 | #39 |
| 3 | Claude Opus 5 | 86.1 | #16 |
| 4 | Grok 4.5 | 86.1 | #54 |
| 5 | Claude Sonnet 5 | 80.6 | #65 |
| 6 | Composer 2.5 | 75 | #130 |
| 7 | Gemini 3.1 Pro (Preview) | 72.2 | #64 |
| 8 | Gemini 3.6 Flash | 72.2 | #57 |
| 9 | Gemini 3.5 Flash | 66.7 | #66 |
| 10 | DeepSeek V4 Flash | 63.9 | #155 |
| 11 | MiMo-V2.5 | 61.1 | #196 |
| 12 | Claude Sonnet 4.6 | 52.8 | #94 |
| 13 | DeepSeek V4 Pro | 52.8 | #112 |
| 14 | Gemini 3.1 Flash Lite | 50 | #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-0-5 · How It Works · Data refreshed daily, snapshot 2026-11-05.