← All benchmarks

WildClawBench

Agents · 2026-05-11

WildClawBench evaluates agents on 60 real-world, long-horizon tasks across productivity, code intelligence, social interaction, search and retrieval, creative synthesis and safety alignment. Each task runs in an isolated Docker container under the OpenClaw harness with browser, shell, file-system, email and calendar tools, and is graded from 0 to 1 by deterministic checks, an LLM judge or both. The score is the unweighted mean of the 60 task scores.

Top models (higher is better)