Disclosure: this article is written and published by DROZlegal, which sells practice-automation software to Ontario law firms. Every claim about OpenAI's product comes from named, dated legal-tech press cited below; we have not tested Astra for Law ourselves, and access is limited at launch. This article is general information, not legal advice, and reflects publicly available information and Ontario law as of October 5, 2026.
What OpenAI launched, in five facts
- What it is. A configuration of OpenAI's GPT-6 Astra model with a legal search index, instructions for legal analysis and writing, and controls aimed at law firms, per LegalTech.ca and LawNext.
- What the index covers. U.S. case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, including the Free Law Project's CourtListener collection, which OpenAI says covers more than 99.9% of published U.S. precedential case law.
- Who gets it first. Firms in a Trusted Access program, through ChatGPT and Codex. API access follows later. LawNext reports the program targets Am Law 200 firms.
- Who built on it. Harvey and Legora are named as API customers, with Thomson Reuters, iManage, Intapp and DeepJudge among plugin developers (LawNext).
- The accuracy figure. OpenAI tested the setup on 200 U.S. questions from the private validation set of Vals AI's Legal Research Bench: 54.0% passed overall correctness, against 38.7% for the same model using web search alone.
A 54.0% pass rate on U.S. legal research questions means 46% of the answers failed the benchmark's overall correctness check. That is OpenAI's own best-case figure, on U.S. law. — from OpenAI's reported Vals AI Legal Research Bench result, as relayed by LawNext
Three questions to put to any U.S.-built legal AI before an Ontario file touches it
The launch answers some of these and leaves others open. This table separates what is published from what you would need to ask.
| Question | What the launch coverage says | What an Ontario firm should ask |
|---|---|---|
| Whose law is in the index? | U.S. only. Neither LegalTech.ca nor LawNext mentions Canadian coverage; LegalTech.ca calls the U.S. focus “an important limitation for Canadian firms.” | Does it return Ontario and federal authorities at all, and if it answers a Canadian question, from what source? An answer built from U.S. precedent or web search is a different thing from a cited Canadian decision. |
| Where does the data go, and what is kept? | OpenAI describes zero data retention on the API for eligible firms in the Trusted Access program, and says ChatGPT Enterprise use is excluded from human review. No data-residency statement was reported. LawNext notes the access agreement runs some 30 pages with exceptions. | Ask for the retention, residency and review terms in writing, for your plan, not the launch headline. Zero data retention is a negotiated arrangement for a defined group of customers; it is not a default for every account of any provider. |
| How often is it wrong, on my kind of question? | 54.0% overall correctness on 200 U.S. research questions at the highest reasoning effort, up from 38.7% with web search alone. | No published figure covers Canadian law. Test it on a closed Ontario matter whose answer you already know, and keep the results in your file. |
The duty that does not move
Ontario's Rule 4.06.1 requires the person signing a factum to certify that they are satisfied as to the authenticity of every authority cited. The rule does not ask which tool found the authority. A citation that came from a model grounded in a U.S. index, or from any model at all, is still yours to confirm against the source.
Our breakdown of the AI disclosure duty in Ontario court filings walks through how the certification works in practice, and the court-by-court tracker shows which Ontario venues add their own AI rules. For what happens when the check is skipped, see the tracker of fabricated AI citations in Canadian courts.
Where Canadian-grounded tools sit
The contrast worth drawing is what a research tool is grounded in. Clio's Vincent now checks citations against more than 567,000 validated Canadian decisions, according to Clio's own announcement; we covered what that does and does not solve in our Vincent post. DROZlegal's research tooling grounds case-law answers in CanLII's verified-citation corpus. See our guide to AI legal research software for Canadian firms for how the options compare, and where popular legal AI tools say they store Canadian firms' data for the residency question.
If Astra for Law reaches Canadian coverage later, the same three questions apply. Until then, treat it as a U.S.-law tool.