OpenAI's Astra for Law: big firms first, right on 54% of questions

On Thursday afternoon, @OpenAI announced Astra for Law, "frontier intelligence built for your practice," a version of GPT-6 Astra packaged with tools, settings, and context for lawyers and the software companies that serve them. About an hour later, @gdb called it "tools and skills to radically accelerate law firms," with "data privacy as a core feature," plus 26 partner-built plugins and 47 community plugins inside ChatGPT.

The announcement drew more than 20,000 likes in its first day, which is a lot of attention for a product most lawyers cannot use yet.

What it actually is

This is not a new model. Jason Boehmig, the former Ironclad CEO who joined OpenAI in June to run its legal products, told LawSites it is a configuration: GPT-6 Astra with instructions for legal writing and analysis, settings tuned for longer and more careful answers, and a new Legal Search Index. The index covers US case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs. The Free Law Project's CourtListener archive supplies the published precedents.

Access goes first to selected firms through a Trusted Access program in ChatGPT and Codex. The named launch partners are Sullivan & Cromwell, which built an agreement analyzer, Ropes & Gray, which built an M&A diligence system, and Cooley, which built an IPO preparation tool. An API version is "coming soon," with no date and no price. Harvey, valued this month at $15.6 billion, and Legora will build on it through that API. So will Thomson Reuters, which is bringing its HighQ matter context into ChatGPT and previewing a CoCounsel connector.

The privacy piece Brockman pointed to is real: zero data retention on the API for eligible firms, and ChatGPT Enterprise usage excluded from human review by default. We covered on Monday why that default matters more than the brochure suggests.

The number OpenAI published

OpenAI ran the full setup 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% of them, against 38.7% for GPT-6 Astra using plain web search. OpenAI calls that a 40% relative improvement. It also found 24% more of the reference cases on case-law questions and retrieved up to 54% more of the target passages from the right opinions.

Read the same figure the other way. The product built for lawyers, with the index and the settings and the maximum thinking budget, still failed the correctness check on nearly half the questions. The test was internal. OpenAI also published a hand-picked pair of examples against Claude, with no explanation of how those two prompts were chosen and no outside audit of either side.

My read

OpenAI is fourth in line. Microsoft shipped a Legal Agent in Word in April, Anthropic launched Claude for Legal in May, and Google announced Gemini Enterprise for Legal in August. Richard Tromans at Artificial Lawyer calls this the battle for centrality: every vendor wants to be the screen the lawyer never leaves, and each one's plugins are a polite way of saying you can visit the others from here. Twenty-six plugins from Thomson Reuters, Clio, and the rest do not change who owns the model, the index, and the model picker.

The 54% is the figure I would show a client. A LexisNexis survey of 543 legal professionals, published September 6, found 83% worried about fabricated information from AI, up from 57% in early 2024. One of the incidents behind that worry was Sullivan & Cromwell apologizing to a New York judge for made-up citations in a filing. The same firm is now a launch partner, and I don't think that's hypocrisy. The tool got better, and the checking stays, which is what the survey respondents were asking for anyway.

One more thing I should own. When Astra launched, my concern was reasoning you cannot read. A legal configuration inherits that. The citations are checkable. The path the model took to reach them is less so.

What this means for a Fox Valley firm

If you run a five-lawyer office in Geneva or Aurora, Trusted Access is not for you this month, and the API has no price. What you can get today is the plumbing around it: Vincent from Clio is already available in Codex, and the LegalZoom, NetDocuments, and CourtListener plugins are live in ChatGPT.

Research is also not where a small firm loses its afternoons. It loses them to the "any news on my case?" call that eats the paralegal's afternoon, and to the consult that no-shows on Thursday. Fixing either needs no 230-million-URL index, only a workflow with a human check at the step where an error would reach a client or a court.

New Face Design's free process audit maps those steps for a small firm and marks where an AI draft is safe to send and where a person signs off. Start here.