OpenAI introduced Astra for Law this past Thursday (the 17th), a configuration of GPT-6 Astra built for people who spend their day reading case files, precedent, and legal opinions. It isn't a new model: it's the same Astra behind ChatGPT, paired with a dedicated legal search index, system instructions tuned for legal writing and analysis, and a set of 26 plugins built around law-firm workflows.
What actually changes
The edge here isn't the model's intelligence, it's what the model can see. The search index covers more than 230 million URLs of US case law, statutes, court rules, and administrative decisions, refreshed daily. In OpenAI's own internal evaluation, Astra for Law passed 54% of an overall-correctness test, against 38.7% for plain Astra using web search alone — 24% more relevant cases found, and 54% more correct passages retrieved from the right opinions.
None of those numbers prove the model "understands" law better. They prove that giving a trained language model better context beats training a bigger one, which is the same lesson AI engineering has been repeating since retrieval became its own discipline: the bottleneck is rarely the model, it's almost always what the model gets to read.
Limited access, for now
Astra for Law is available only to selected firms through the Trusted Access program, inside ChatGPT and Codex — API access is "coming soon," with no date attached. Early clients include Latham & Watkins, Ropes & Gray, Cooley, and Sullivan & Cromwell: firms that already had the budget and in-house legal staff to absorb the risk of an AI getting a citation wrong in a filing.
That caution answers a known problem: language models hallucinate case law with embarrassing confidence, and more than one US lawyer has already been sanctioned for citing a case that never existed. A dedicated search index cuts that risk; it doesn't eliminate it, and to its credit, OpenAI doesn't promise zero errors anywhere in the announcement.
Why this matters if you build AI products
Astra for Law is a textbook case for shipping a vertical AI product without retraining anything: take a frontier model, attach a well-curated domain data source, tune the system prompt, and measure the lift against a baseline. OpenAI already showed this instinct elsewhere, as we covered when the company formalized how it discloses its own models' misalignment incidents; here, the same appetite for process shows up on the commercial side.
Worth asking: how many verticals like this one does OpenAI plan to package before it becomes, in practice, a holding company for legal, financial, and medical products running on the same model underneath?