OpenAI launches Astra for Law with a 230-million-document legal index and a 54% benchmark pass rate

OpenAI introduced Astra for Law on September 17: a configuration of its GPT-6 Astra model wrapped in a dedicated legal search index -- more than 230 million URLs of US case law, statutes, regulations and court rules, refreshed daily -- plus instructions tuned for legal research and writing. It is not a new model; it is GPT-6 Astra pointed at a purpose-built corpus, offered first to law firms and the software vendors that sell to them.

Access rolls out in two tracks. Harvey, Legora and Thomson Reuters get API access to build their own products on top of it; a smaller group of Am Law 200 firms gets a Trusted Access program inside ChatGPT and Codex directly, with zero data retention on the API and enterprise usage excluded from OpenAI's human review by default. OpenAI is working with law firm Latham & Watkins specifically on the permissions layer -- ethical walls, client instructions, firm-level oversight -- that a tool touching privileged material needs before a firm will run it on a live matter.

Legal research is a deliberate proving ground, not an easy first vertical. Billable-hour economics mean firms will pay well for genuine accuracy gains, but the same economics mean a wrong citation reaches a judge, not just a customer-support ticket -- malpractice exposure is real and immediate in a way it isn't for most chatbot use cases. That's the case for building the permissions layer first: a tool this easy to misuse in a filing has to earn trust on confidentiality before it gets judged on capability at all.

On Vals AI's Legal Research Bench -- 200 validation questions the firm keeps private specifically so vendors can't train against it -- Astra for Law passed the overall correctness check on 54% of questions, against 38.7% for GPT-6 Astra using plain web search: a 15.3-point gain, or about 40% better in relative terms. On case-law questions specifically, it surfaced 24% more relevant case references than the web-search baseline.

Read the other direction, 54% correctness means Astra for Law still didn't fully pass the correctness check on 46% of the benchmark's own validation questions -- a real improvement over ad hoc web search, not a solved problem. (Vals AI's private, rotating question set is a deliberate anti-gaming design -- a public benchmark invites a vendor to quietly optimize for the test rather than the underlying task.) Vals AI keeps its methodology and question set private, which means the 54% figure can't currently be checked against the underlying questions by anyone outside the benchmark's own operator and OpenAI.

The stakes for that remaining 46% aren't abstract. As of June 9, 2026, a tracker maintained by Damien Charlotin at HEC Paris had logged 1,598 court cases worldwide where a judge found or clearly implied a party had relied on hallucinated AI-generated citations or quotes -- adding roughly 8 new cases a day. Penalties have escalated well past embarrassment: one federal case closed in December 2025 with about $109,700 in combined sanctions and fees, and a federal judge went further in June 2026, canceling a trial outright and suspending both lead attorneys from the district for two years. None of those cases involve Astra for Law, which didn't exist yet when they were filed -- but they're the backdrop any legal-AI tool launches into now, and the likely reason OpenAI's Trusted Access program leads with confidentiality controls and firm oversight rather than raw capability.

Astra for Law doesn't compete with Harvey or Legora so much as sit underneath them -- both are named API customers, not rivals being disintermediated, and Thomson Reuters is previewing its own CoCounsel connector into the same index rather than building a competing one. The bet OpenAI is making is that the legal industry's bottleneck was never model quality alone; it was a search index good enough, and access controls strict enough, that a firm would trust either one with a real client matter.

That framing also explains why OpenAI shipped this as an index and a permissions layer rather than a smarter model: the 15-point benchmark gain came entirely from better retrieval over a better corpus, using the same underlying GPT-6 Astra the web-search baseline also ran on. The next real jump for legal AI, on this evidence, is more likely to come from whoever builds the next better index than from whoever trains the next bigger model.