OpenAI introduced Astra for Law yesterday. It isn’t a new model. It’s GPT-6 Astra with three things wrapped around it: a search index for US law, fixed instructions for legal analysis and writing, and controls for confidential client work.
An index instead of a web search
The index covers US case law, statutes, regulations, court rules and administrative decisions — more than 230 million URLs by OpenAI’s count, with sources added daily. The case-law collection comes from the Free Law Project, the non-profit behind CourtListener, which says it covers more than 99.9 percent of published US precedential case law.
That’s the actual news here. The foundation isn’t coming from a publisher, it’s coming from a public-interest database. OpenAI is careful to say the index complements licensed content from providers such as Thomson Reuters rather than replacing it. You can read that as courtesy or as notice.
54 against 38.7 percent
The test ran on 200 questions from the private validation set of Vals AI’s Legal Research Bench. At the highest reasoning effort, Astra for Law passed the overall correctness check on 54.0 percent of questions; GPT-6 Astra with web search alone managed 38.7 percent. On case-law questions the legal configuration found 24 percent more reference cases and up to 54 percent more relevant passages from the right opinions.
Respectable jumps. But mostly they say something about web search: a lawyer hunting for authority doesn’t need a tool that rummages around the open internet. They need one that knows which decision binds.
Legora sits on both sides
Select law firms and API customers get first access, with Harvey and Legora named. On top of that come 26 new ecosystem plugins connecting ChatGPT to tools firms already use, like Relativity and Clio.
Legora is a familiar name. The company showed up two days ago as a customer in Anthropic’s Salesforce announcement. Legal tech vendors clearly aren’t tying themselves to one lab — they’re tying themselves to whichever model works best, and both labs are happy to name the same firms as proof.
The model isn’t the product anymore
What’s interesting about Astra for Law is what it isn’t. No training run, no fine-tune, no separate set of weights. An index, instructions, permissions — and OpenAI says these capabilities will carry forward to future models.
It’s the same shift Anthropic is running with skills and connectors: value moves out of the model and into everything arranged around it. For users that means something uncomfortable. When you ask which tool is better in a given field, comparing base-model benchmarks no longer answers it. You have to compare the configuration, and from the outside that’s nearly impossible to check.
Sources: