A configuration, not a new model
OpenAI released Astra for Law on 17 September: its flagship GPT-6 Astra with settings, tools and context assembled for legal research and drafting. No new weights, no new architecture — the product is the scaffolding around the model.
The centrepiece is a dedicated legal search index covering more than 230 million URLs of US case law, statutes, regulations, court rules and administrative decisions. OpenAI built it in part from materials published by CourtListener, run by the non-profit Free Law Project.
What OpenAI says it is worth
The company tested the configuration on 200 legal research questions drawn from Vals AI’s Legal Research Bench. At its highest reasoning effort, Astra for Law passed the overall correctness check on 54.0% of questions, against 38.7% for GPT-6 Astra using ordinary web search — a 40% relative improvement. On case-law questions specifically, OpenAI reports finding 24% more reference cases and retrieving up to 54% more relevant passages from opinions.
Those are OpenAI’s own figures on a third-party benchmark, published by the vendor rather than measured by the benchmark’s maintainers. The gap between 54% and the remaining 46% is also the part that matters most in practice: a legal research tool that is wrong on nearly half of a hard question set is an assistant, not an authority.
Who can use it
Astra for Law goes first to selected US law firms through a Trusted Access programme aimed at Am Law 200 firms, with zero data retention and exclusion from human review. API customers including Harvey and Legora will be able to build on it. OpenAI has not published pricing.
Around it sits an ecosystem layer: 26 vendor plugins, 9 community plugins and 47 custom skills, connecting ChatGPT to tools firms already run, including Relativity, Clio and iManage. Thomson Reuters is named among the partners. Firms listed as having worked on testing and design include Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins and Wachtell Lipton.
The competitive read
The companies best placed to be disrupted by this are the ones OpenAI has signed up. Harvey, valued at $15.5bn in a round that closed this month, sells legal AI built on frontier models; Thomson Reuters sells the research databases the index competes with. Both are partners here, which is how OpenAI has handled every vertical so far — ship the layer, keep the incumbents inside the tent.
What to watch
Whether other verticals follow on the same pattern, and whether Vals AI or another independent evaluator reproduces the 54% figure. A legal research claim is unusually checkable, because the cases either exist and say what the model says they say, or they do not.
