THE SIGNAL IN ONE SENTENCE
OpenAI has introduced Astra for Law, a legal version of GPT-6 Astra connected to a search index for United States case law, statutes, regulations, court rules, and administrative decisions. The index spans more than 230 million URLs and incorporates Free Law Project's collection, which the nonprofit says covers more than 99.9 percent of published United States precedential case law. OpenAI also adds instructions for legal analysis and writing, selected-firm access through ChatGPT and Codex, data controls for confidential work, and connections to specialist legal products. That is a meaningful change from asking a general chatbot to roam the web and improvise a memo. It is not a digital partner who can safely sign her name and go home. On OpenAI's private 200-question validation set, the complete Astra for Law setup passed an overall correctness check on 54.0 percent of questions. The comparison system, GPT-6 Astra with ordinary web search, passed 38.7 percent. A 40 percent relative improvement is useful. A 54 percent pass rate also leaves a lot of red ink. The plain signal is that better retrieval can move legal AI from confident wandering toward inspectable research, but only if firms preserve the chain from question to authority, passage, jurisdiction, current treatment, client fact, lawyer review, and accountable advice.
01
WHAT ACTUALLY CHANGED
OpenAI announced Astra for Law on September 17 as a foundation for law firms and legal technology companies. It combines GPT-6 Astra with a legal search index and instructions for research, analysis, and writing. Selected firms receive initial access through a Trusted Access program in ChatGPT and Codex. OpenAI says the API model, gpt-6-astra-law, is coming soon, so API availability should not be treated as generally live today.
The index can search United States case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily. OpenAI says its work with Free Law Project brings CourtListener's case-law collection into the experience. Free Law Project says that collection encompasses more than 99.9 percent of published United States precedential case law and uses court scrapers for continuing updates. Coverage does not by itself prove that every source is current, controlling, complete, or correctly applied to a client's facts.
OpenAI tested the complete setup on 200 questions from a private validation set of Vals AI's Legal Research Bench. At the highest reasoning effort, Astra for Law passed the benchmark's overall correctness check on 54.0 percent, compared with 38.7 percent for GPT-6 Astra using web search alone. That is a 15.3 percentage-point gain and about a 40 percent relative improvement. Vals describes its broader benchmark as realistic United States research across eight practice areas with lawyer-authored, peer-reviewed rubrics, but the exact 200 questions and OpenAI run are not publicly inspectable.
OpenAI reports that, on case-law-focused questions, Astra for Law found 24 percent more reference cases than the web-search comparison. On an audited set of target passages, it retrieved up to 54 percent more relevant passages from the correct opinions at the same reasoning effort. More authorities and passages can help a lawyer see the field. They can also produce a larger pile of plausible material to validate, distinguish, update, and fit to the correct court, date, procedural posture, and client problem.
The launch adds operational controls and integrations. Eligible firms can receive zero data retention on the API, while ChatGPT Enterprise use is excluded from human review by default. OpenAI says it is working with Latham & Watkins on permissions, ethical walls, client instructions, and firm oversight. It also announced 26 partner-built plugins, nine community plugins with 47 adaptable skills, and general availability of ChatGPT for Word. Those are product and design commitments, not evidence that every firm configuration already enforces every client restriction.
02
WHY THIS MATTERS
Legal research fails in ways that look respectable. A fabricated citation is obvious once someone checks it. A real case quoted accurately can still be useless because another court controls, the decision was reversed, a statute changed, the passage was dicta, or the facts point the other way. OpenAI's own launch example says another frontier model returned a holding reversed on appeal. Astra found a closer case in that prompt. The example is illustrative company material, not an independent head-to-head verdict.
Retrieval changes the economics of verification. When a system can surface a relevant opinion and the exact passage, the lawyer can spend less time guessing search terms and more time testing authority, negative treatment, factual fit, and strategy. That is the helpful version. The dangerous version converts saved search time into a larger volume of unchecked output because the answer now arrives with convincing links.
Confidentiality is a system, not a toggle. Zero data retention addresses one storage path. A firm must still govern who may send which matter data, what a plugin can retrieve, where prompts and outputs are logged, how ethical walls propagate, whether a vendor or subcontractor can access content, how exports enter the document system, and what happens when a client forbids a use. The label matters less than the complete data route.
A private benchmark can guide development and cannot carry the entire trust claim. Readers cannot inspect the 200 questions, scoring disagreements, source freshness, practice-area balance, error severity, repeated-run variance, cost, latency, or how many apparently correct answers contained a dangerous extra statement. A missed citation in a research exercise and a wrong deadline in a live matter are both failures and not remotely the same risk.
The market is shifting from one chatbot to a layered legal stack: frontier model, search index, firm data, specialist product, document editor, plugin, access policy, review workflow, and human sign-off. Each layer can improve the answer and each adds another permission boundary, version, vendor, log, and failure mode. Firms need a matter-level record that shows which system did what and which lawyer accepted the result.
03
WHERE IT COULD HELP
- Require every research answer to preserve the question, jurisdiction, date cutoff, retrieved authorities, exact passages, subsequent-history check, negative treatment, source links, omitted adverse authority, model and tool versions, reviewer, corrections, and final disposition in the matter file
- Validate the system on a firm-owned set of completed matters and deliberately difficult questions, scoring controlling authority, factual fit, current treatment, adverse authority, unsupported propositions, confidentiality, cost, latency, and whether a lawyer can reproduce the answer from primary sources
- Build matter-level permission maps before connecting plugins: identify the client instruction, responsible lawyer, authorized team, approved sources, prohibited systems, ethical wall, retention rule, geographic restriction, export destination, incident route, and person who can revoke access
- Separate draft assistance from legal judgment by marking model-generated propositions, assigning a qualified reviewer, blocking unverified citations from final work product, preserving substantive edits, and making the approving lawyer responsible for the complete document rather than a sample
- Measure outcomes beyond speed, including research corrections, missed authority, client objections, write-offs, privilege or confidentiality incidents, junior-lawyer training, accessibility, cost passed to clients, and whether the tool expands useful service or merely increases document volume
KEEP A HAND ON THE WHEEL
Astra for Law is a company-announced product in selected access, not a generally available autonomous lawyer. The 54.0 percent result comes from OpenAI's run on 200 questions from a private Vals AI validation set. It is a benchmark pass rate, not proof of general legal accuracy, professional competence, client outcomes, privilege protection, deadline reliability, or safe use without review. The 40 percent figure is a relative improvement from 38.7 percent, while the absolute gain is 15.3 percentage points. The 24 percent and up-to-54 percent retrieval improvements apply to narrower case-law comparisons described by OpenAI. More results are not automatically better results. Free Law Project's greater-than-99.9-percent claim concerns published precedential case-law coverage, not every legal source, unpublished decision, current treatment, jurisdictional rule, client fact, or licensed commentary. Zero data retention applies to eligible API use, and ChatGPT Enterprise is excluded from human review by default, but firms still control identities, plugins, exports, matter permissions, ethical walls, client instructions, and local records. API availability is described as coming soon. Watch for public benchmark details, independent replication, error categories, repeated-run stability, negative-treatment testing, jurisdictional coverage, complete data-flow documentation, client consent practices, incident reporting, pricing, and evidence from live supervised work.
04
TERMS WORTH KEEPING
OPEN GLOSSARY CARD
Legal search index
A structured collection that lets a system locate legal sources and passages using metadata, text, links, and other retrieval signals.
OPEN GLOSSARY CARD
Controlling authority
Law that a court or decision maker must follow for the relevant issue, jurisdiction, and procedural setting.
OPEN GLOSSARY CARD
Citation validation
Checking that a cited source exists, says what the writer claims, remains current, applies in the relevant jurisdiction, and supports the proposition in context.
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on September 18, 2026.
PUBLICATION RECEIPT: Revision 1. Published September 18, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US