OpenAI published Introducing Astra for Law. The offering is GPT-6 Astra with a legal search index and instructions for legal analysis and writing. Harvey and Legora will be able to build on it through the API. In ChatGPT it shows up as GPT-6 Astra Law. Early access is selected law firms through Trusted Access in ChatGPT and Codex. The help page adds that this is initially United States firms, for eligible lawyers and people working under their supervision. The API is coming later.
Legal research, they write, often begins with finding the exact right authority, locating the relevant passages, and understanding how relevant and binding they are. The index is one of the tools Astra for Law can use. It searches U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily. They credit Free Law Project and CourtListener for a collection covering more than 99.9 percent of published U.S. precedential case law. They say this sits alongside licensed products such as Thomson Reuters, not in place of them.
They tested the full setup on 200 U.S. legal research questions from the private validation set of Vals AI's Legal Research Bench. At the highest reasoning effort for both systems, Astra for Law passed the evaluation's overall correctness check on 54.0 percent of questions, against 38.7 percent for GPT-6 Astra using web search alone. They call that a 40 percent relative improvement. On case-law questions it found 24 percent more reference cases. On an audited set of target passages, at the same reasoning effort, it retrieved up to 54 percent more relevant passages from the correct court opinions.
Custom instructions then apply that research to the client's facts, develop arguments or deal terms, and flag weaknesses. That can mean telling a holding from other remarks in an opinion, or showing how a contract exception moves risk. The lawyer can examine the authorities herself. The help page is blunter: review the answers and cited sources before relying on them.
They also describe firm-built workflows on ChatGPT Enterprise: Sullivan & Cromwell on agreement review, Ropes & Gray on a data room, Cooley on IPO filings. Plugins connect ChatGPT to tools firms already use. The product is a talking model with a better library, sold as support for a lawyer's judgment.
Editorial
The hunt they named is how a junior used to get paid and how a junior used to get trained. Find the case. Pull the passage. First-pass the stack. Draft from the file. Firms billed those hours. The person doing them learned, slowly, which holding travels and which cite is the wrong court. Astra for Law is that pile with a model on top.
Keep their 54 percent. A better index beats raw web search. It still fails the overall correctness check on almost half the set, on their bench, at max reasoning. The leftover work is catching that half. A graduate whose only offer is a longer first pass than GPT-6 Astra is already competing with a machine that does not need a seat.
I do not have a headcount forecast. This is selected U.S. firms, not every legal team. If the apprenticeship was “grind research until judgment appears,” the grind is the part they automated. Judgment still has to come from somewhere. Firms that cut the grind and keep only the partner will get cheaper drafts and more expensive mistakes. The student who can show what the index missed is the hire.