OpenAI puts GPT-6 Astra inside law firms with a 230M-URL legal index
Two directions today: the frontier labs are building the plumbing for one profession's daily work, and a consumer agent is taking over the family calendar.
OpenAI launched Astra for Law, a configuration of GPT-6 Astra connected to a legal search index covering more than 230 million URLs, and began offering it to selected firms through a new Trusted Access program aimed at the Am Law 200. OpenAI is explicit that this is not a new model but a configuration — domain instructions for legal analysis and writing, response-length settings, and legal-specific tooling it says will expand over time. The index is the substance: case law, statutes, regulations, court rules and administrative decisions, drawing on CourtListener's collection from the nonprofit Free Law Project, which covers more than 99.9% of published US precedential case law. Access ships with zero data retention on the API and ChatGPT Enterprise usage excluded from human review by default, and OpenAI says it is working with Latham & Watkins on information permissions, ethical walls and firm oversight. It appears in the model picker as "GPT-6 Astra Law" and, when it reaches the API, as gpt-6-astra-law.
The benchmark OpenAI published is both the pitch and the caveat. On the 200-question validation set of Vals AI's Legal Research Bench, at the highest reasoning effort for both systems, Astra for Law passed the overall correctness check on 54.0% of questions against 38.7% for GPT-6 Astra using web search alone — a 40% relative improvement, on questions where the baseline got most answers wrong. It found 24% more reference cases and retrieved up to 54% more relevant passages from the correct court opinions, and its answers run roughly twice as long. Longer and better-cited is not the same as correct, and 54% on a research benchmark leaves a lot of room; the honest reading is that the configuration fixes a weak baseline rather than making research reliable enough to file without a lawyer.
Why it matters: OpenAI is playing the platform card in a vertical its own partners and its rivals already occupy. Anthropic shipped a legal configuration of Claude in May with more than 20 connectors, Google launched Gemini Enterprise for Legal in August — we covered the opening moves then — Google launches Gemini Enterprise for Legal to take on Claude in law — and Harvey and Thomson Reuters now train models of their own. OpenAI's answer is to own the index and the model while letting the specialists plug in: 26 partner-built plugins from Thomson Reuters, iManage, Intapp, DeepJudge and others, nine community plugins and 47 skills, with Sullivan & Cromwell, Ropes & Gray and Cooley showing tools their own lawyers built. The access-to-justice question got a nod rather than an answer — OpenAI called it paramount and said there would be more to share later.
Google Labs turned CC, its experimental agent, into a family agent — one with its own Google account, up to six members, and a permissions model that decides what it sees of each person's inbox. School emails, sports schedules, bills and vet appointments arrive across separate accounts, so CC sorts them into a shared "Your Day Ahead" brief, adds calendar events and tasks, and can pre-fill fillable PDFs such as permission slips, waivers and activity registrations, asking for anything missing. It cannot sign or submit a form, enter payment details or buy anything, and it cannot process video or audio. Each member chooses which senders to share; the weekly suggestions of new senders are generated privately, and Google says CC does not store or learn from those emails. US users aged 18 and over on a personal account can upgrade or join the waitlist. CC stays a Labs experiment separate from Gemini Apps and Workspace — and once a member accepts an invitation they cannot be removed individually, only by deleting the whole group.
Household logistics is the context-rich, low-stakes ground where a consumer agent can actually earn trust, and the published limits are the interesting part: an agent that will fill in your child's permission slip but refuses to sign it is drawing a deliberate line, not shipping half a feature.
Andrew Ng told Bloomberg TV that renewed warnings of human extinction from AI are "much more science fiction than science," and that the latest "wave of PR" is probably intended to shape regulation. Ng, who helped start Google Brain and Coursera, made the same argument at a US Senate forum in 2023 — that large companies amplify fear in order to write rules that pull up the ladder behind them. What has changed is that his two categories now sit in one law: the EU AI Act's list of systemic risks names both the engineering failure modes he wants prioritised and the loss-of-control risks he dismisses in the same recital, with duties that bite above a compute threshold, landing on the biggest developers and nobody else. His framing is contested from the other end of the same argument — the researchers who left the frontier labs this month describe voluntary disclosure as the only functioning mechanism: The researchers who left say AI transparency is voluntary — Trump says he has no extinction concerns. Both can hold at once. A warning can be a PR wave and the labs can still be the wrong sole authors of the rules.
What to watch: whether Astra for Law's numbers get an independent re-run the way its predecessor's did — OpenAI quietly rewrote Astra's benchmark numbers after launch — and whether any of the firms in the Trusted Access program publish what their lawyers actually use it for.
If a legal AI gets half the research questions wrong, should the law firm or the vendor carry the liability when a filing cites a case that doesn't hold? Tell us in the comments.
