OpenAI has launched Astra for Law, a legal-industry edition of its GPT-6 Astra model, complete with a purpose-built legal search index, confidentiality controls and a partner list that reads like a legal awards dinner. It is a full-stack play for a single profession, and it arrives with numbers attached. Per OpenAI's announcement, dated 17 September, the full setup passed 54.0% of questions on the validation set of Vals AI's Legal Research Bench, against 38.7% for GPT-6 Astra with web search alone. Those are OpenAI's own tests, and they frame the story's real question: can a frontier model earn the trust of the one profession that checks footnotes for a living?
OpenAI Watch analysis
What happened
Astra for Law combines GPT-6 Astra with a legal search index covering U.S. case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with sources added daily. A partnership with the Free Law Project, the non-profit behind CourtListener, brings case law covering more than 99.9% of published U.S. precedential decisions into the research experience, per the announcement.
Access starts narrow. The offering goes first to selected law firms through a Trusted Access programme in ChatGPT and Codex, with an API version, gpt-6-astra-law, to follow. OpenAI says eligible firms get Zero Data Retention on the API and ChatGPT Enterprise usage excluded from human review by default, with Latham & Watkins advising on information permissions, ethical walls and firm oversight.
The firms are not window dressing. Per the announcement, Sullivan & Cromwell built an agreement analyzer that turns its negotiating playbooks into proposed redlines, Ropes & Gray built a deal diligence system that traces findings back to source in the data room, and Cooley built GO Public, which carries deal changes across an IPO filing. Harvey and Legora will build Astra for Law into their own products through the API.
Our top picks
- A legal search index spanning 230 million URLs
U.S. case law, statutes, regulations and court rules, refreshed daily, so answers can cite real authorities. - Trusted Access confidentiality controls
Zero Data Retention on the API and Enterprise usage excluded from human review by default, aimed squarely at client-confidence duties. - 26 partner plugins and 9 community plugins
iManage, Intapp, DeepJudge and Thomson Reuters connectors, plus 47 custom skills from legal engineers. - ChatGPT for Word is generally available
Proofreading, suggested edits and formatting flags inside the tool lawyers already draft in. - Firm-specific builds from forward-deployed engineers
Sullivan & Cromwell's agreement analyzer, Ropes & Gray's diligence system and Cooley's GO Public IPO tool. - Harvey and Legora on the API
The legal tech layer gets to wrap its own products around the foundation.
Why it matters
Law is the vertical everyone insists AI will transform and the one least forgiving of a confident wrong citation. A product that leads with grounding, authority-tracking and confidentiality controls is an admission that trust, not raw intelligence, signs the purchase order. If the numbers survive independent scrutiny, research-heavy legal work gets faster and cheaper, and the billable hour's moat gets a proper inspection.
The cast list matters as much as the benchmark. Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins and Wachtell do not lend their names to science projects, and their quotes describe working tools rather than pilots. When the most conservative buyers in the professional world start building on a frontier model, every other white-collar industry takes notes.
Our read
This is OpenAI growing up in public: less 'look what the model can do', more 'here are the ethical walls'. The 54.0% figure is a company result on a private validation set, so treat it as a claim awaiting independent replication rather than a settled score. Even discounted, the direction is unmistakable: the competitive frontier has moved from chat to configured, governed deployments, and law is the beachhead with the deepest pockets. If you work in or around legal services, the useful move is asking which of your systems will connect, and who owns the audit trail when the model cites the wrong case.
What to watch
- API general availability
of gpt-6-astra-law, and its pricing. - Independent runs
of the Vals AI Legal Research Bench by anyone other than OpenAI. - Harvey and Legora shipping
Astra for Law features into customer-facing products. - Regulator and bar guidance
as Trusted Access firms formalise AI use in privileged work.
Discussion spark: OpenAI excludes these firms' Enterprise usage from human review by default to protect client confidences. Is that the right trade, or does privileged work deserve more oversight than ordinary use, not less?
Sources and evidence
- Introducing Astra for Law – openai.com (17 September 2026, 20:15 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.