OpenAI's Astra for Law pairs GPT-6 Astra with a 230-million-URL legal index, open to selected law firms only (2026)

OpenAI introduced Astra for Law on September 17, 2026: GPT-6 Astra, its latest model, configured with a legal search index, instructions for legal analysis and writing, and a set of legal plug-ins. The product appears in the model picker as "GPT-6 Astra Law" and in the API as gpt-6-astra-law, and OpenAI is offering it first to selected law firms through a Trusted Access Program, in ChatGPT and Codex. API access is "coming soon," with no date. 1

What launched

Signal	Confirmed detail	Action window
A configuration of an existing model	GPT-6 Astra plus the legal search index and legal instructions; the base model is otherwise unchanged. 1	Existing GPT-6 Astra work continues as it is.
Legal search index	More than 230 million URLs of U.S. case law, statutes, regulations, court rules and administrative decisions, with sources added daily. Free Law Project's CourtListener collection, which covers more than 99.9% of published U.S. precedential case law, is part of it. 1 2	Check the covered sources against your jurisdictions.
Access ahead of the API	Selected firms via Trusted Access in ChatGPT and Codex; Harvey and Legora will build on it as API customers. 1	There is no self-serve sign-up, and OpenAI has published no pricing.
Data controls	Zero data retention on the API for eligible firms, ChatGPT Enterprise usage excluded from human review by default, and governance design work with Latham & Watkins. 1 3	Confirm which plan inherits which control before a client matter goes near it.
Ecosystem	26 partner plug-ins, including iManage, Intapp, DeepJudge, Thomson Reuters HighQ, Relativity and Clio, plus 9 community plug-ins and 47 skills; ChatGPT for Word shipped generally available the same day. 1	Check whether the tools your firm runs are on the list.

What the numbers say

OpenAI measured the complete configuration 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% of questions, against 38.7% for GPT-6 Astra using web search alone — a 40% relative improvement. OpenAI also reports 24% more reference cases on case-law questions and up to 54% more relevant passages from the correct opinions. The comparison runs against OpenAI's own base configuration on a held-out split, and the side-by-side example OpenAI published with it used a prompt the company picked, in which Claude Fable 5.1 returned a holding reversed on appeal and Astra for Law returned two matching precedents. 1

Vals AI's own public leaderboard for the same benchmark, which grades a question correct only when every item in a practitioner-written rubric passes, puts the leading general models at 55.29% in a three-way tie across 27 models. That is a different question split from OpenAI's, so the two figures are not a head-to-head; together they set the scale, at roughly half of answers still missing a required element. The benchmark's hardest category is reconciling conflicting authority, at 20.7% all-pass against 28.4% overall. 4

Why it matters

OpenAI is selling a vertical rather than a model: a frontier model, a case-law corpus supplied through a nonprofit and its partners, and a governance program, taken to law firms directly while Harvey, Legora and Thomson Reuters keep building on the same models. Access is this week's constraint. A firm outside the Trusted Access Program has nothing to install, and builders have no API date to plan around. The change worth watching is whether a licensed index and practitioner-written instructions move a frontier model on a task the public leaderboard still scores as unsolved.