Infrathon Benchmark Tests LLM Agents on Real-World AI Infra Optimization; Claude Opus 5.5 Leads

RulinShao · x · 2026-11-17

Researchers from UW, Allen AI, Stanford and others launched Infrathon, a benchmark of 25 real-world, open-ended AI infrastructure optimization tasks with 12–24-hour budgets (150 agent runs, 10,000+ GPU-hours). Agents edit systems, pass a Controller code review, and are scored by a Judge on held-out workloads. Claude Opus 5.5 leads all six models with a mean rank score of 91.2, winning 17/25 tasks — including boosting mean KV-cache compression from 6.82× to 19.86× while improving video quality.

Original post →

More from Infra

Infra channel →