ClawBench: New Agent Benchmark Shows Frontier Models Score Only 33% on Real-World Everyday Web Tasks

arXiv·high signal

Researchers published ClawBench, a benchmark of 153 everyday online tasks across 15 life categories and 144 live production websites — not sandboxes. Claude Sonnet 4.6 scored 33.3% and GPT-5.4 scored just 6.5%, compared to 65-75% these same models achieve on traditional web benchmarks. The gap reveals that real-world write-heavy tasks (form filling, multi-step checkout, appointment booking) remain dramatically harder than the read-heavy tasks that existing benchmarks measure. A lightweight interception layer blocks final submissions for safe evaluation.