← All benchmarks

Legal Agent Benchmark - Anthropic public 1235-task set

Professional Work

Public Harvey LAB with 16 defective tasks excluded before evaluation from 1251, leaving 1235. Anthropic's internal harness exposes bash and Python while preserving Sonnet 4.6 grading and all-pass criteria, and reports the mean over five trials. This public cohort has its own factor group and must not be pooled with Harvey held-out or AA private panels.

Top models (higher is better)