OpenAI is selling its most capable model to the legal profession. On September 17 the company released Astra for Law, a product that pairs GPT-6 Astra, its strongest system, with a legal retrieval index and specialized instructions for analysis and writing. The package is aimed at law firms and legal technology companies, two markets where the value of an AI that cites real cases is highest.
The retrieval library is the heart of the product. It covers more than 230 million URLs of American case law, statutes, court rules, and administrative decisions, a corpus intended to anchor the model’s answers in actual legal authority rather than its memory. The case law portion comes from CourtListener, run by the Free Law Project, and covers more than 99.9 percent of published U.S. precedent, the company said.
The distinction between searching and remembering matters in law. A general-purpose model can produce a plausible-sounding citation that does not exist, a failure mode lawyers call a hallucination and cannot tolerate. Astra for Law is designed to retrieve from the index and cite from it, giving the model a factual ground for its analysis that a purely generative system lacks.
OpenAI offered numbers to make the case. On a set of 200 legal research questions prepared by Vals AI, Astra for Law answered correctly 54 percent of the time. GPT-6 Astra with only web search, without the legal index, scored 38.7 percent. The gap is the argument for the product: the specialized retrieval layer, not the underlying model, is what turns a general assistant into a legal tool.
The product is also a platform. Companies like Harvey and Legora, which build software on top of OpenAI’s models through its API, can use Astra for Law as a foundation for their own products. That positions OpenAI as a supplier of legal infrastructure rather than a direct competitor to the startups it serves, at least for now.
The legal market has been one of the most contested in enterprise AI. Law firms spend heavily on research and drafting, and the promise of automating even a fraction of that work has drawn a crowd of startups and incumbents. OpenAI’s entry with its strongest model raises the stakes, particularly given the breadth of the index it has assembled.
The accuracy numbers are honest about the limits. A 54 percent overall accuracy rate means the product gets a significant share of legal research questions wrong, and OpenAI did not claim otherwise. For the firms that will pay for it, the question is whether a tool that is right slightly more than half the time on hard questions, and presumably far more often on routine ones, is worth the price and the review burden.
Legal professionals have been cautious adopters. The cost of an error in a brief or a filing can be severe, and firms have governance around any tool that touches client work. Astra for Law’s retrieval-first design is aimed at that caution, giving lawyers a way to check where an answer came from rather than trusting the model’s word.
CourtListener’s role in the product is more than a data source. The Free Law Project, which runs it, is a nonprofit that has spent years building a free, comprehensive archive of American law, and its inclusion gives Astra for Law a citation base that is public and verifiable. For a product sold to lawyers, that provenance matters: the citations the model produces can be checked against a source the profession already trusts.
The legal AI market is among the most crowded in the industry. Startups like Harvey and Legora have raised substantial capital on the promise of automating research and drafting, and the large publishers that dominate legal information, such as the owners of Westlaw and LexisNexis, are building their own AI tools. OpenAI’s entry with its strongest model and a broad index raises the competition, and its decision to sell through the startups rather than only around them blurs the line between partner and rival.
The accuracy figures are a candid admission of limits. Being right 54 percent of the time on a set of hard research questions is a real capability and a real shortfall, and OpenAI presented both. The firms that adopt the product will build review processes around its output, treating it as a first draft and a citation finder rather than a source of finished legal work. That is the honest position of the tool today, and the company did not pretend otherwise.
The release extends a pattern in OpenAI’s product strategy: take the strongest general model and wrap it in domain-specific retrieval and instructions for each industry it wants to sell into. Law is among the highest-value targets, and the combination of the 230-million-URL index with the Astra model is OpenAI’s opening bid for a share of the billions of dollars the legal industry spends on research each year.
