On September 17, 2026, OpenAI shipped Astra for Law, a version of GPT‑6 Astra configured specifically for legal work. It pairs the model with a searchable index of U.S. case law, statutes and regulations, plus instructions that teach it how lawyers actually reason about authority.
It is the most serious move OpenAI has made into the legal market. It is also, for now, almost impossible to get: access is limited to selected law firms through an invitation-only program, and OpenAI has published no pricing.
Here is what Astra for Law actually is, what its benchmark numbers do and do not prove, who can use it today, and what to do if you are one of the many lawyers who cannot.
Key Takeaways
- Astra for Law is a configuration, not a new model. It combines GPT‑6 Astra with a legal search index and custom legal instructions.
- The index covers more than 230 million URLs of U.S. case law, statutes, regulations, court rules and administrative decisions, with sources added daily.
- It passed the overall correctness check on 54.0% of benchmark questions, against 38.7% for GPT‑6 Astra with plain web search. That is a real gain. Note that this is an all-or-nothing pass rate, not an error rate: it means nearly half the answers carried at least one flaw the graders caught.
- Access is gated. Selected firms only, via OpenAI's Trusted Access Program, inside ChatGPT and Codex. API access is "coming soon" with no date.
- No pricing has been announced, and the index covers U.S. law only.
- Solo practitioners, small firms and non-U.S. lawyers have no path to it yet, which is where self-serve tools like LegesGPT remain the practical option.
What Is Astra for Law?
Astra for Law is not a new model. OpenAI describes it as a foundation for firms and legal tech companies to build on, and it has three layers stacked on top of GPT‑6 Astra, which OpenAI calls its latest and most powerful model.
| Layer | What it does |
|---|---|
| GPT‑6 Astra | The underlying frontier model doing the reasoning |
| Legal search index | A retrieval tool covering U.S. primary law, updated daily |
| Legal instructions | Guidance on legal analysis and writing conventions |
That third layer is the part most people overlook, and it is arguably the most interesting. According to OpenAI, the instructions push the model to distinguish a court's holding from its passing observations, to address cases that weaken an argument instead of ignoring them, and to explain how a contract exception shifts risk between parties.
Those are exactly the habits that separate a usable research memo from a confident-sounding summary. A general-purpose chatbot will happily cite dicta as though it were binding. That failure mode is a large part of why general models have been risky for research, and it is the gap a purpose-built GPT built for law is meant to close.
When it reaches you, it appears in the ChatGPT model picker as "GPT‑6 Astra Law" and in the API as gpt-6-astra-law.
The Legal Search Index: 230 Million URLs of U.S. Law
The retrieval layer is the substance of the announcement. Astra for Law can search U.S. case law, statutes, regulations, court rules and administrative decisions across a corpus of more than 230 million URLs, with new sources added daily.
The most consequential piece is a collaboration with the Free Law Project, the nonprofit behind CourtListener. Its case-law collection covers more than 99.9% of published U.S. precedential case law, and that collection now sits inside the research experience.
Two things are worth noting about how OpenAI framed this.
First, it positioned the index as a complement to licensed content rather than a replacement, naming Thomson Reuters directly. Second, Thomson Reuters is a launch partner, not a casualty: it is bringing HighQ matter context into ChatGPT and previewing a forthcoming CoCounsel Legal connector. Anyone reading this as a Westlaw killer is reading it wrong.
The limit is geographic. This is a U.S. corpus. If your practice touches English, Canadian, Australian or EU law, OpenAI describes no coverage of it.
What the Benchmark Actually Says
OpenAI tested the full Astra for Law setup against 200 U.S. legal research questions drawn from the private validation set of Vals AI's Legal Research Bench. The benchmark scores whether the model finds the relevant sources and passages, and whether its answers meet defined evaluation criteria.
Comparing both systems at their highest reasoning effort, except where noted:
| Metric | Astra for Law | GPT‑6 Astra + web search |
|---|---|---|
| Overall correctness check passed | 54.0% | 38.7% |
| Reference cases found (case-law questions) | 24% more | baseline |
| Relevant passages from correct opinions (at equal reasoning effort) | up to 54% more | baseline |
OpenAI characterizes the headline result as a 40% relative improvement. On its own terms, that is a substantial jump, and the retrieval gains are the clearest evidence that grounding a model in real primary law beats letting it browse the open web.
Now read the same number the other way. On OpenAI's own benchmark, 46% of Astra for Law's answers did not clear the overall correctness check. Read that precisely, because it is easy to overstate. The check is all or nothing: OpenAI labels it an "all-pass rate," and an answer that misses a single evaluation criterion fails it. Alongside it OpenAI plots a "weighted rubric score" on which both systems land in the high 70s to low 90s. So 46% is not an error rate. It is the share of answers carrying at least one defect the graders flagged, which on a legal research test, at maximum reasoning effort, is still close to half the output needing a fix that a lawyer has to find.
Four caveats belong with those figures:
- The baseline is OpenAI's own model with web search, not Westlaw, Lexis, or any competing legal research product. It shows the configuration works. It does not rank Astra for Law against the tools firms already pay for.
- The numbers are self-reported on a private validation set, with no independent reproduction published.
- The two "54%" figures are unrelated. One is an answer-level pass rate, the other is a passage-retrieval improvement. They are easy to conflate and they measure different things.
- The 54.0% figure is an all-pass rate, not an accuracy score. OpenAI reports a separate weighted rubric score on which both systems sit much higher. Quoting 54% as "how often it is right" overstates the failure, and quoting the rubric score as "how good it is" understates it.
OpenAI's launch post also included a head-to-head example against Claude Fable 5.1, in which Claude returned a holding that had been reversed on appeal in one scenario and reported finding no matching case in another. Treat that as vendor marketing rather than evidence: it is a hand-picked pair of prompts with no disclosed selection method and no independent judge. If you want a fairer read on how general models handle legal work, our Claude vs ChatGPT for lawyers comparison covers it directly.
The practical takeaway is unchanged by any of this. Verification is still your job, and no benchmark number transfers that duty to the vendor.
Find on-point cases in plain English
Describe your issue and LegesGPT surfaces relevant case law and statutes, with direct links to the source so you can verify every result.
Search case lawWho Can Actually Use Astra for Law?
This is where enthusiasm meets reality. Astra for Law is being offered initially to selected law firms through OpenAI's Trusted Access Program, inside ChatGPT and Codex. Eligible firms can extend it to lawyers and to people working under their supervision, for professional legal work.
What that means in practice:
- Not generally available. You cannot sign up and start using it.
- API access is "coming soon" with no published date, though Harvey and Legora will be able to build on it.
- No pricing at all has been announced. Not per seat, not per token. OpenAI's benchmark charts do plot model cost per answer, and the configuration is visibly more expensive to run than plain web search, but no price for the product itself appears anywhere.
- U.S. law only in the index, with no stated plan for other jurisdictions.
- ChatGPT Enterprise is the delivery vehicle for the firm-specific work, which implies enterprise procurement.
So the population that can use Astra for Law today is a small number of firms OpenAI has selected. The named participants (Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins, Wachtell) are all elite large firms, though OpenAI states no size requirement. Everyone else is waiting, including the solos, the two-to-fifty-attorney firms, and the in-house teams that make up most of the profession.
Trust, Confidentiality and Governance
The governance work may matter more to firms than the benchmark does, because confidentiality is the objection that actually blocks AI adoption in legal practice.
For eligible firms, the Trusted Access offering includes Zero Data Retention on the API, and ChatGPT Enterprise usage is excluded from human review by default. Those two commitments address the question every general counsel asks first: where does client information go, and who sees it.
OpenAI is also working with Latham & Watkins on the harder institutional problem: designing information permissions, ethical walls, client instructions and firm oversight. Ethical walls are not a nice-to-have. A firm that cannot demonstrate a conflicted matter team was walled off from another team's confidential information has a professional responsibility problem, not a software problem. Building that into the AI layer is a precondition for real deployment, and it is the kind of control that any serious legal AI for law firms has to account for.
What Firms Actually Built on It
Three named firms shipped internal tools with help from OpenAI's forward-deployed engineers, adapting ChatGPT Enterprise with custom interfaces and connections to proprietary data.
- Sullivan & Cromwell built an agreement analyzer that brings the firm's negotiating playbooks and selected precedents into review of a new deal. It surfaces risks that only appear when provisions are read together, then turns findings into proposed redlines and draft client advice.
- Ropes & Gray built a deal diligence system modeled on how its lawyers work through a data room. It traces findings back to the source document and flags questions that could affect an acquisition, such as whether key customer contracts require notice or consent.
- Cooley built GO Public, which carries its capital markets expertise from drafting the IPO filing through identifying risks management needs to see. When deal terms change, it propagates the change across the filing.
Wachtell, Lipton, Rosen & Katz is doing more exploratory work with OpenAI on how AI can support sophisticated legal judgment.
Read those carefully and a pattern emerges: each is a bespoke build created with OpenAI's forward-deployed engineers, wired into proprietary data the firm already owns. None is a feature you can switch on. The moat here is not the model, it is the firm's own work product plus the engineering time to wire it in.
The ecosystem story is broader. OpenAI launched 26 partner-built plugins connecting ChatGPT to tools firms already use, including Relativity, Clio, iManage, Intapp and DeepJudge, plus 9 community plugins carrying 47 custom skills from practitioners at LegalQuants, LECG and Skills.law. ChatGPT for Word also went generally available, so drafting-stage proofreading and suggested edits now live where the document does.
What This Means If You Are Not at a Large Firm
Strip away the launch coverage and the situation for most lawyers is simple. The best legal research configuration OpenAI has ever built is real, it is measurably better than a chatbot with web search, and you cannot buy it.
That gap is worth naming honestly, because it is the whole story for solo practitioners, small firms, in-house teams and anyone practicing outside the United States.
LegesGPT covers that ground today, self-serve and without a procurement cycle. Where Astra for Law is a foundation that firms build on, LegesGPT is a finished workflow you log into:
- Answers legal questions with verified citations and direct links to the source, so checking authority is a click rather than a separate research session
- Case law and statute search in plain language, across 38+ jurisdictions rather than U.S. law alone
- Document review that identifies risks, flags problematic clauses and proposes concrete changes
- AI drafting for contracts and legal documents, with 100+ attorney-drafted templates
- E-signature to sign and send what you drafted, in the same subscription
Pricing is public, which is its own kind of answer to an unpriced enterprise product: Basic at $19.99/month, Plus at $49.99/month including 50 document reviews, and Premium at $99.99/month with unlimited review and priority support. Annual billing saves about 30%, and there is a 3-day trial for $1 if you want to test it against your own matters first. Full details are on the plans and pricing page.
Two honest limitations, since a tool with no drawbacks is a sales pitch. LegesGPT is web-only: there is no public API, no mobile app and no Word add-in, so if drafting inside Word is central to your workflow, OpenAI just shipped something LegesGPT does not match. And a self-serve product does not give you forward-deployed engineers to wire your firm's precedent bank into a custom tool, which is precisely what Sullivan & Cromwell and Cooley got.
For a wider view of what else is available, our roundup of AI legal research software options and the best Harvey AI alternatives both cover the field beyond OpenAI.
Get legal answers with verified citations
Ask any legal question and get a clear answer grounded in real sources — every citation links back to the statute or case it came from.
Ask a legal questionHow to Use Any AI for Legal Research Without Getting Burned
That 46% all-pass shortfall is the reason this section exists. It applies to Astra for Law, and it applies more forcefully to every general-purpose model you might use instead.
- Open every citation. Not the model's summary of the case, the case. A citation that looks plausible and formats correctly can still point at nothing.
- Confirm it is still good law. Reversed, vacated and superseded authority is the single most dangerous output, and it is exactly what OpenAI's own comparison example caught a competing model doing.
- Read the passage the model relied on. Retrieval improvements mean better passages are being found. They do not mean the passage says what the model claims it says.
- Check the holding against the dicta. If the tool cannot tell you which is which, you have to.
- Treat the output as a research starting point, not a filed document. It narrows where you look. It does not replace looking.
- Know your jurisdiction's disclosure rules. Courts have been sanctioning lawyers over fabricated citations since 2023, and many judges now have standing orders requiring counsel to disclose AI use, certify that a human verified every citation, or both. Check your judge's standing order before you file, not after.
None of this is a reason to avoid AI research. It is a reason to keep a lawyer in the loop, which is what our guide on how to effectively use AI for legal research walks through in more depth.
Frequently Asked Questions
What is OpenAI's Astra for Law?
Astra for Law is a configuration of GPT-6 Astra built for legal work, announced on September 17, 2026. It combines the model with a legal search index covering more than 230 million URLs of U.S. case law, statutes, regulations, court rules and administrative decisions, plus custom instructions for legal analysis and writing. It appears in ChatGPT as "GPT-6 Astra Law" and in the API as gpt-6-astra-law.
Is Astra for Law powered by GPT-6?
Yes. Astra for Law runs on GPT-6 Astra, which OpenAI describes as its latest and most powerful model. Astra for Law is not a separate model but that model plus a legal retrieval tool and legal instructions layered on top.
How accurate is Astra for Law?
On 200 U.S. legal research questions from the private validation set of Vals AI's Legal Research Bench, Astra for Law passed the overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone. OpenAI calls that a 40% relative improvement. The check is all-or-nothing (OpenAI labels it an all-pass rate), so the 46% that did not clear it is the share of answers with at least one flagged defect rather than an error rate. Every citation still needs to be verified by a lawyer.
Who can use Astra for Law and what does it cost?
Initially only selected law firms admitted to OpenAI's Trusted Access Program, which extends access to their lawyers and to people working under those lawyers' supervision, inside ChatGPT and Codex. API access is described as coming soon, with Harvey and Legora among the companies that will build on it. OpenAI has announced no pricing, and the legal search index covers U.S. law only.
What can small firms use instead of Astra for Law?
Because Astra for Law is gated to firms OpenAI selects, and unpriced, solos and small firms need a self-serve option. LegesGPT answers legal questions with verified citations, searches case law and statutes across 38+ jurisdictions, reviews documents for risk, drafts contracts and handles e-signature, starting at $19.99/month with a 3-day trial for $1.
