OpenAI launches Astra for Law with a 230-million-URL legal research index

OpenAI introduced Astra for Law on September 17, a version of its GPT-6 Astra model paired with a legal search index spanning more than 230 million URLs of U.S. case law, statutes, regulations, court rules, and administrative decisions. The product targets law firms and legal technology companies directly, entering a market already served by established players Thomson Reuters CoCounsel and Harvey.

Where the data comes from

Much of the underlying case law comes from the Free Law Project, the nonprofit organization behind CourtListener, which OpenAI says covers more than 99.9% of published U.S. precedential case law. Sources are added to the index daily. This matters for a legal research tool specifically because legal citations require currency — a case that was good law last month can be overturned or distinguished by a ruling this week, and a research tool built on a stale index risks producing confidently wrong citations.

The benchmark number that matters more than the marketing

On a published 200-question set from the Legal Research Bench, Astra for Law achieved 54% overall correctness — a 40% relative improvement over GPT-6 Astra using generic web search alone, which scored 38.7% on the same questions. That's a meaningful jump, and OpenAI deserves some credit for publishing the number rather than only promoting the improvement percentage. But a 54% correctness rate on a dedicated legal research benchmark is also a genuine limitation worth stating plainly: for a profession where a single fabricated or outdated citation in a court filing can result in sanctions, a tool that gets roughly half of research questions fully right is a research aid requiring verification, not a replacement for a lawyer checking the underlying source.

Early access and competitive position

Initial access runs through a "Trusted Access" program inside ChatGPT and Codex, with API access planned for later. Confirmed early-adopter firms include Latham & Watkins, Ropes & Gray, Cooley, and Sullivan & Cromwell — four Am Law 100 firms, which signals OpenAI is pursuing large firms with existing budget for legal AI tools rather than starting with smaller practices. The product also ships with 26 partner-built and 47 community-created plugins for expanded functionality, an ecosystem play similar to how OpenAI has built out ChatGPT's broader plugin store.

Harvey's own leadership was quoted acknowledging that Astra for Law "demonstrated impressive research depth and sensitivity to authority" — a notably candid response from a direct competitor, suggesting the benchmark improvement is real rather than inflated marketing.

What this means for the legal AI market

OpenAI entering legal research directly, rather than leaving the vertical entirely to specialized startups like Harvey or incumbents like Thomson Reuters, changes the competitive calculus for smaller legal AI vendors: they now compete against a frontier lab with a general-purpose model and a purpose-built legal index, not just against each other. For law firms evaluating these tools, the 54% correctness figure is the number to press vendors on directly — ask what "correctness" means on their internal benchmarks, and whether verification workflows are built into the product or left entirely to the associate reviewing the output.