Skip to content
Advertisement
OpenAI’s Astra for Law 6.2 tops Harvey LAB-AA v1.1 with 20.8% score

OpenAI’s Astra for Law 6.2 tops Harvey LAB-AA v1.1 with 20.8% score

The legal-focused model beat its own predecessor and rival systems on a hallucination-gated benchmark of 120 real-world legal tasks

OpenAI’s Astra for Law 6.2 now leads Harvey’s LAB-AA v1.1 benchmark with a hallucination-gated all-pass rate of 20.8%. That puts it ahead of every other model in the evaluation, including OpenAI’s own GPT-6.2 Astra.

What the benchmark actually measured

Harvey LAB-AA v1.1 is built from 120 private tasks drawn from real-world legal work. Those tasks span 24 different practice areas, so a model cannot coast on mastery of contracts alone.

Under those rules, Astra for Law 6.2 posted 20.8%. Its predecessor, GPT-6.2 Astra, recorded 15.0%, a gap of 5.8 percentage points. Muse Wonder Full trailed further back at 13.1%.

The more telling number may be the error rate. Astra for Law 6.2 logged a material hallucination rate of just 0.03 per task, the lowest of any model in the evaluation.

Put plainly, the model produced about three material fabrications for every hundred tasks.

Advertisement

Who gets access, and how

The upgrade, announced around November 17, 2026, also widens the door. ChatGPT Enterprise customers can reach Astra for Law through OpenAI’s Trusted Access Program.

OpenAI also plans to make the model available via API to partners, including Harvey and Legora.

Harvey has welcomed the change, pointing to improved quality in brief-writing.

Julio Pereyra, Harvey’s Head of Legal Research, said Astra for Law 6.2 produces briefs that judges prefer over those from its predecessor approximately 90% of the time.

The model has also been climbing on other leaderboards. In a Vals AI Legal Research Benchmark evaluation conducted on November 16, 2026, Astra for Law reached a full-credit score of 53.8% on research questions.

Back in September, that figure sat at nearly 46%. A gain of around eight points in roughly two months suggests the improvements are not confined to a single test.

What this means for law firms and legal tech vendors

The partner arrangement complicates the picture in an interesting way. Harvey designed the benchmark, praised the upgrade and is set to receive API access to the model that topped it.

The absolute numbers still deserve a sober reading. A 20.8% all-pass rate means the best model in the field failed to fully clear roughly four out of every five tasks under strict scoring.

The judge-preference figure cited by Pereyra points in the same direction. Better briefs are valuable, but the comparison is against a predecessor model, not against seasoned attorneys.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.