OpenAI Debuts Astra for Law With 54% Research Accuracy

OpenAI has introduced Astra for Law, a specialized legal configuration of its GPT-6 Astra model designed for law firms and legal technology developers. The system integrates a massive U.S. legal search index containing more than 230 million URLs, updated daily. In partnership with the nonprofit Free Law Project, this index incorporates CourtListener's database, which covers over 99.9% of published U.S. precedential case law. On the Vals AI Legal Research Bench, Astra for Law achieved a 54.0% correctness rate, representing a notable improvement over the 38.7% scored by standard GPT-6 Astra using web search at its highest reasoning setting. The system also retrieved up to 54% more relevant passages from correct judicial opinions, though its 46.0% failure rate on the benchmark means human review remains essential.

Several prominent law firms have already collaborated with OpenAI engineers to build custom workflow tools. Sullivan & Cromwell developed an agreement analyzer to apply firm playbooks and draft client advice, while Ropes & Gray built a deal-diligence system to track contract requirements. Cooley created GO Public to assist with IPO preparation, and Wachtell, Lipton, Rosen & Katz is working on tools for complex legal analysis. Latham & Watkins is advising on governance and ethical walls. Additionally, the launch features 26 partner plugins and nine community plugins covering 47 custom skills. These integrations connect ChatGPT to platforms like Thomson Reuters, iManage, Relativity, Clio, Intapp, and DeepJudge, while startups Harvey—which has raised over $1 billion—and Legora plan to build on the system via its upcoming API.

To address strict confidentiality requirements, OpenAI is offering a Trusted Access program that supports Zero Data Retention on the API and excludes ChatGPT Enterprise usage from human review. Alongside the launch, ChatGPT for Word has become generally available. However, OpenAI has not yet disclosed a release date, pricing, eligibility criteria, or usage limits for the planned API. Because the high reasoning settings required for thorough legal research consume more compute and increase latency, firms will need to carefully evaluate the ultimate operational costs of the system.