OpenAI says Astra for Law leads legal intelligence benchmarks
OpenAI has released an updated Astra for Law, a GPT-6.2 Astra configuration that the company says leads several legal intelligence evaluations. ChatGPT Enterprise customers can now request access through the Trusted Access Program, including through the API.
- Astra for Law is a GPT-6.2 Astra configuration with an expanded legal search index.
- It earned full credit on 20.8% of LAB-AA assignments, ahead of GPT-6.2 Astra and Muse Wonder Full.
- OpenAI reports 0.03 material hallucinations per assignment in the latest LAB-AA comparison.
- Full-credit answers on Vals AI’s Legal Research Bench rose to 53.8% from about 46%.
- ChatGPT Enterprise customers can request access through the Trusted Access Program and API.
A legal configuration built on GPT-6.2 Astra
Astra for Law combines GPT-6.2 Astra with an expanded legal search index and instructions for researching questions, weighing authorities and writing answers. OpenAI introduced the configuration to a small group of firms in September, including Cooley, Ropes & Gray and Sullivan & Cromwell.
Since then, the company says its research team has expanded the legal search index by over a third and made improvements to legal reasoning. The index now spans over 310 million URLs, after the addition of over 80 million new URLs, and covers more than 99.9% of published U.S. precedential case law. It also includes material from international jurisdictions such as the UK, Australia and Canada.
OpenAI says the updated configuration earned full credit on LAB-AA tests nearly 40% more often than GPT-6.2 Astra, the runner-up in that comparison.
Top results on legal agent and brief-writing tests
Artificial Analysis’s LAB-AA evaluates whether a model can complete legal assignments from start to finish. The test includes 120 assignments across 24 areas of law, including drafting deposition outlines, reviewing commercial leases and strategizing ownership changes in acquisitions. To receive full credit, an answer must meet every requirement without making a material claim unsupported by the source documents.
Astra for Law ranked first, earning full credit on 20.8% of assignments. GPT-6.2 Astra scored 15.0%, while Muse Wonder Full scored 13.1%. OpenAI says Astra for Law achieved its result with 54% less output per assignment than Muse Wonder Full, averaging about 64,000 output tokens compared with 139,000.
The company also says Astra for Law tied with GPT-6 for the lowest rate of material hallucinations in the latest LAB-AA comparison, at 0.03 per assignment. Claude Opus 5.5 recorded 0.59 material hallucinations per assignment. Astra for Law had zero material hallucinations on tasks that met the all-pass criteria.
On LAB: USCL, a benchmark co-developed by Harvey and Grubel for researching and writing briefs in real U.S. federal cases, Astra for Law earned one of the top scores in both test versions. When models researched the law themselves, Astra for Law scored 1,844 compared with 1,817 for Opus 5.5. When given the record and precedents cited by the original lawyer, the scores were effectively even at 1,883 and 1,882.
Research performance improves, while access expands
OpenAI reports that Astra for Law’s performance also improved on Vals AI’s Legal Research Bench. In the company’s September evaluation, it earned full credit on about 46% of questions. In Vals AI’s November 16 evaluation, it received full credit on 112 of 208 questions, or 53.8%, and earned 90.9% of the available points including partial credit.
OpenAI describes the roughly eight percentage point increase in full-credit answers in under two months as evidence of progress in legal research, a core focus of the updated release. The company says it continually refines Astra for Law’s settings and instructions in response to practitioner feedback, while also improving tools such as the legal search index.
All ChatGPT Enterprise customers can request access through the Trusted Access Program, including through the API. The release gives legal teams a model configuration designed specifically around legal research, authority evaluation and long-form legal writing, while the benchmark results remain claims from OpenAI and the organizations that conducted the evaluations.
OpenAI: Astra for Law leads in legal intelligence

