InfraBench - Mean Effective Score: leaderboard
InfraBench (arXiv 2608.11234v2, Table 3): agent CLIs paired with pinned models operate infrastructure environments on 12 tasks spanning hardware, operating systems and distributed systems across deployment, maintenance, repair and decommissioning, three attempts each. Difficulty-weighted mean verifier score, 0-100; higher is better. Each row is a CLI plus model configuration.
Metric: Difficulty-weighted mean effective score (0-100; mean of three attempts). Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 | 89.2 | #20 |
| 2 | Grok 4.5 | 84.3 | #54 |
| 3 | Claude Opus 5 | 83.1 | #16 |
| 4 | Claude Opus 4.8 | 80.2 | #39 |
| 5 | Claude Sonnet 5 | 77.2 | #65 |
| 6 | Gemini 3.6 Flash | 75.5 | #57 |
| 7 | Composer 2.5 | 72.6 | #130 |
| 8 | Gemini 3.5 Flash | 72.1 | #66 |
| 9 | Gemini 3.1 Pro (Preview) | 70.9 | #64 |
| 10 | DeepSeek V4 Flash | 60.6 | #155 |
| 11 | MiMo-V2.5 | 53 | #196 |
| 12 | Claude Sonnet 4.6 | 52.6 | #94 |
| 13 | Gemini 3.1 Flash Lite | 46.1 | #200 |
| 14 | DeepSeek V4 Pro | 45.5 | #112 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=infrabench-mean-effective-score · How It Works · Data refreshed daily, snapshot 2026-11-05.