OpenAI’s Legal Model Posts 20.8% All-Pass Rate on Harvey Test
Astra for Law 6.2 led 120 legal tasks across 24 practice areas, but met the benchmark’s full-credit standard on about one in five.

OpenAI’s Astra for Law 6.2 led Harvey’s 120-task legal benchmark with a 20.8% all-pass rate, a result that puts it ahead of tested rivals while showing most tasks did not earn full credit.
Harvey LAB-AA v1.1 covers 24 areas of legal work. GPT-6.2 Astra posted a 15% all-pass rate, while Muse Wonder Full scored 13.1%, leaving Astra for Law 6.2 ahead of its predecessor by 5.8 percentage points.
Astra for Law 6.2 also had the evaluation’s lowest material hallucination rate, at 0.03 per task. The all-pass result means the model met the benchmark’s full-credit standard on about one in five tasks.
Separately, Astra for Law scored a full-credit 53.8% on research questions in the Vals AI Legal Research Benchmark on Nov. 16. The score was nearly 46% in September.
ChatGPT Enterprise customers can access Astra for Law through OpenAI’s Trusted Access Program. OpenAI plans to offer API access to partners including Harvey and Legora. The results follow OpenAI’s Nov. 17 launch of GPT-6.2 Astra and GPT-6.2 Sol.