OpenAI has launched Astra for Law, a legal configuration of GPT-6 Astra. The headline numbers are striking, but the details matter: its search index covers more than 230 million URLs of U.S. legal material, not 230 million documents stored inside the model, and its privacy controls differ depending on how a firm accesses it.

Astra for Law is not a claim that GPT-6 Astra has memorised 230 million legal documents. OpenAI says the product pairs GPT-6 Astra with a Legal Search Index that searches more than 230 million URLs covering U.S. case law, statutes, regulations, court rules and administrative decisions, with new sources added daily. The point is to help legal users move from the facts of a matter to relevant authorities and passages they can inspect themselves. That distinction matters because the source retrieval layer is separate from the model's own generated answer.

The privacy controls are more specific than a blanket promise of zero retention. OpenAI says eligible firms can use Zero Data Retention on the API. ChatGPT Enterprise usage is handled differently and is excluded from human review by default. OpenAI also says enterprise customer data is not used to train its models unless the customer explicitly opts in. Firms evaluating the product therefore need to check which deployment route they are using and which retention, review and access settings apply to that route.

Astra for Law is initially being offered to selected U.S. law firms through Trusted Access in ChatGPT and Codex, with API access coming later. That limited rollout means there is not yet broad real-world evidence about how consistently the legal configuration performs across different firms, practice areas and document sets. OpenAI tells users to review answers and cited sources before relying on them, which is especially important in legal work where an incorrect authority or an overstated conclusion can have material consequences.

If the search layer reliably surfaces relevant authorities and the model can apply them accurately to a client's facts, the immediate value is likely to be faster research, comparison and drafting rather than autonomous legal judgment. The useful question for a firm is not whether the system can produce a polished answer. It is whether lawyers can trace that answer back to the right authorities, understand the uncertainty and fit the tool inside the firm's confidentiality and approval processes.

The next evidence to watch is practical: citation accuracy, coverage of the relevant authorities, how often lawyers have to correct the model, and whether the privacy controls work cleanly inside existing firm systems. Astra for Law is a significant product launch, but the strongest claims about its value will need to come from measured use rather than the launch announcement itself.

For legal teams, the practical test is straightforward: can the system surface the right authority, preserve the required confidentiality controls and make its work easy to verify? Those are the measures worth watching as selected firms begin using it. I'm Isla, and that is the state of play for now.