OpenAI has packaged a legal-specific version of its flagship model, calling it Astra for Law. The offering pairs GPT-6 Astra with a dedicated legal search index and instructions tuned for legal writing and analysis. That index reaches administrative rulings, court procedures, regulatory filings, statutes and case law, spanning more than 230 million URLs that OpenAI says are updated daily.
Harvey and Legora, two of the best-funded legal AI startups, are named as API customers that will build their own products on top of Astra for Law, not as rivals OpenAI is displacing. Still, the practical effect is that OpenAI is now selling directly to the same law firms those startups court, even while it remains their infrastructure supplier. That dual role, supplier and competitor, is OpenAI’s structural bet on the category, not a claim the company has made in its own words.
The search index leans on a partnership with Free Law Project, the nonprofit that runs CourtListener, whose case-law collection reaches more than 99.9 percent of all precedential rulings published in the United States. That data sits alongside licensed sources firms already use from providers such as Thomson Reuters, rather than replacing them.
On performance, OpenAI ran its own comparison. Pulling 200 questions from a benchmark called Vals AI’s Legal Research Bench, specifically that benchmark’s private validation set, the company tested Astra for Law’s full setup at the highest reasoning effort. It passed the correctness check on 54 percent of questions, compared with 38.7 percent when GPT-6 Astra relied on ordinary web search instead, a 40 percent relative gain by OpenAI’s own accounting.
On case-law questions specifically, the legal configuration surfaced 24 percent more reference cases than the web-search baseline. On an audited subset, it surfaced up to 54 percent more of the relevant passages, pulled from correct court opinions. These are OpenAI’s figures, tested against a benchmark OpenAI chose, and no third party has replicated them yet.
Access starts narrow. Astra for Law rolls out first through a Trusted Access Program inside ChatGPT and Codex, with API availability described as coming later. It will show up in the model picker as “GPT-6 Astra Law” and in the API as gpt-6-astra-law. For firms in the program, OpenAI is offering Zero Data Retention on the API and excluding ChatGPT Enterprise usage from human review by default, working with Latham & Watkins on the governance design around information walls and client instructions.
OpenAI’s forward-deployed engineers have also built bespoke tools with specific firms: an agreement analyzer for Sullivan & Cromwell, a deal-diligence system for Ropes & Gray, and an IPO-filing tool called GO Public for Cooley. Alongside those, OpenAI is launching 26 partner plugins connecting ChatGPT to tools firms already run, including Relativity, Clio, iManage and Intapp, plus 9 community plugins built by practicing lawyers and legal engineers.
For Thomson Reuters and other incumbent legal research providers, the real risk is not the plugin they built for this launch. It is that OpenAI now owns a case law index sitting inside the model itself instead of behind a separate subscription. Every firm evaluating its legal AI budget next quarter will need to decide what it still has to license separately.
Based on OpenAI’s official announcement of Astra for Law, published on openai.com.
