OpenAI has introduced Astra for Law, a GPT-6 Astra configuration aimed at legal research. It is not a separate foundation model trained from scratch. The product combines the general model with a legal-source index, specialized instructions and tools that find cases and return the passages supporting an answer.
OpenAI says the index contains more than 230 million URLs. It includes CourtListener, the Free Law Project database covering more than 99.9% of published precedential US federal case law. Breadth does not make every result controlling authority. Jurisdiction, date and a case's later treatment remain essential.
The improvement is real, but comes from a private test
On a 200-question evaluation prepared by Vals AI, Astra for Law reached 54% overall correctness versus 38.7% for GPT-6 Astra with web search. Against that same baseline, it retrieved 24% more reference cases and as much as 54% more relevant passages, according to OpenAI's published results. OpenAI presents both as relative improvements rather than percentage-point gains.
That is a meaningful difference, but not enough reliability for unsupervised action. Even the higher score leaves many questions answered incorrectly. This is also not an open independent comparison that anyone can reproduce today. Our GPT-6 Astra analysis explains how tools, environment and budget can affect a benchmark as much as the underlying model.
The responsible conclusion is narrower: connecting a capable model to a curated legal corpus improves retrieval over sending it onto the general web. It does not prove that the system can replace professional interpretation, citator checks or the accountability of the lawyer signing the work.
Who can use it and how data is handled
Astra for Law is currently available to selected firms through Trusted Access. OpenAI has announced an API model named gpt-6-astra-law, but says only that it is coming soon. The product also connects to legal plugins and skills. Those integrations can shorten a workflow while adding their own permissions and contractual terms.
OpenAI says business conversations are excluded from model training by default, and that eligible API organizations can use zero data retention. Those controls do not remove the need to review contracts, data residency, legal privilege and third-party access before uploading a matter.
The launch fits OpenAI's broader move toward connected agents and tools. In legal work, the valuable output is not an answer that sounds certain. It is a shorter path to sources that a professional can open, validate and cite with their context intact.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?