PinchBench v2

Overview

PinchBench, from Kilo Code, measures how well large language models work as the brain of an OpenClaw agent on real tasks such as scheduling, research, writing, coding, email and file management. Version 2 has 147 tasks across 11 categories, each graded by automated checks, a Claude Opus rubric judge, or both. The leaderboard accepts public submissions and a model can have several runs; this board records the average score across all published runs rather than the single best run. Its tasks differ from the original PinchBench, so scores are not comparable.