INITIAL RESULTS ON LEGAL AGENT BENCHMARK

Summary from the TLDR AI newsletter

Harvey baselined frontier models on its Legal Agent Benchmark holdout under an "all-pass" standard requiring every rubric criterion to pass, and Claude Opus 4.7 led at just 7.1% with Sonnet 4.6 at 5.4%, Opus 4.6 at 4.2%, GPT-5.5 at 2.1%, and Gemini 3.5 Flash at 0.8%, signaling legal work is far from saturated by frontier intelligence.