On September 17, 2026, OpenAI announced "Astra for Law," a product for legal practice. It combines GPT-6 Astra with legal search and instructions for analysis and drafting, and is offered as a foundation that law firms and legal-tech companies can use to build their own work-focused AI with their own expertise. In OpenAI's evaluation, the answer pass rate on US legal research questions improved over the same base model using web search alone. However, the initial rollout is limited to selected firms. Alongside the performance figures, it is worth checking which laws can be searched, who can use the product, and how client information is handled.
Combining legal search with specialized instructions
Astra for Law's legal search index covers more than 230 million URLs. It can search US case law, statutes, and administrative regulations, as well as court rules and agency decisions, and materials are added daily. The index includes a collaboration with the Free Law Project, the nonprofit that runs the CourtListener case law database. However, the URL count cannot be read as a count of court decisions.
In legal research, after finding documents that seem relevant, one must read which passages support a conclusion and consider whether they apply to the matter at hand. To connect search and analysis, OpenAI also built in legal-specific instructions. The aim is to apply what has been researched to a client's circumstances, build arguments and contract terms, and point out weaknesses and uncertainties. Where to search and how to construct answers are configured together for legal practice. OpenAI's announcement also explains that the legal search complements the paid content and products of specialist providers.
Examples of embedding firm-specific knowledge were also shown. Sullivan & Cromwell is building an analysis tool on ChatGPT Enterprise that uses its negotiation approaches and selected precedents to review new contracts. It looks for risks that arise when clauses are combined and leads to drafts of proposed revisions and client advice, on the premise that lawyers scrutinize and revise the content. Ropes & Gray's acquisition due diligence support and Cooley's IPO preparation tool "GO Public" are likewise examples tailored to each firm's way of working.
Connections to existing business software are also expanding. OpenAI announced 26 plugins made by partner companies, citing iManage, which handles case files, and DeepJudge, which uses past matters for comparison, among others. There are nine community-made plugins built by lawyers and others, including 47 custom skills. Thomson Reuters connects HighQ matter information to ChatGPT and has previewed a future connector for CoCounsel Legal. Connections already available need to be distinguished from those only announced in advance.
A 15.3-point improvement, measured against the same base model
The evaluation OpenAI published on September 17 used 200 US legal research questions from a private validation set of Vals AI's "Legal Research Bench." With reasoning effort set to maximum for both configurations, the share of answers passing the overall correctness evaluation was 54.0% for Astra for Law and 38.7% for GPT-6 Astra using web search alone.
Calculating from OpenAI's published figures, the difference in pass rates is 15.3 points, and the relative improvement is about 39.5%. The difference is 54.0 − 38.7, and the relative improvement is (54.0 − 38.7) ÷ 38.7 × 100, rounded to one decimal place. The "40% relative improvement" the company cites corresponds to this level; it does not mean the pass rate rose by 40 points. This is a recalculation from published values, not an independent replication, and it does not use unrounded figures.
That the comparison is against the same base model also matters. It shows that changing the whole configuration, including specialized search and instructions, improved how answers were rated. However, it is not an experiment isolating the contribution of the legal search index alone, nor does it show superiority over every competing legal product.
The announcement also gives search-result gains separate from answer pass rates. At maximum reasoning effort, on case-law-focused questions, it reportedly found 24% more cited cases. In an audited set of target passages, at the same reasoning effort, it says relevant passages retrieved from the correct decisions increased by up to 54%. The latter 54% is a relative increase in retrieved material, with a different denominator and evaluation target from the 54.0% answer pass rate.
Understanding how the evaluation works also changes how the pass rate should be read. According to Vals AI's explanation, the primary metric counts an answer as passing only if it satisfies all of a question's required criteria, separate from metrics that allow partial credit. Grading is done by a large language model. So it cannot be interpreted to mean that every non-passing answer is entirely wrong, nor that the pass rate directly reflects the accuracy of individual citations.
Moreover, Vals AI's public ranking is based on results from a 208-question test set. Because the question set differs from the 200-question validation set OpenAI used, Astra for Law's 54.0% cannot be placed alongside the public ranking figures to determine its position. What can be compared is, first, the difference between the two configurations measured under the same conditions. Whether the same improvement would be obtained in work involving Japanese law cannot be judged from this US-law evaluation.
Starting with selected US firms; data handling differs by access route
Initial availability is through Trusted Access in ChatGPT and Codex. The official help page describes the eligible users as selected US law firms, with use by qualified lawyers and people working under their supervision. The name shown on the model selection screen is "GPT-6 Astra Law."
The API is planned for later, with the identifier gpt-6-astra-law. OpenAI names Harvey and Legora as API customers, but SiliconANGLE reported that no specific availability date or pricing was given. Treating it as a product ordinary users can choose right away could lead to mistaken adoption plans.
Client-information handling should likewise be read by separating eligibility from the conditions of each service. Lining up the scope shown in OpenAI's announcement and official help gives the following.
| Item to check | Published scope or measure |
|---|---|
| Initial eligibility | Selected US law firms; qualified lawyers and people under their supervision |
| Scope of legal search | US case law, statutes, etc.; performance under Japanese law was not part of this evaluation |
| Data handling in the API | Includes zero data retention (ZDR) for eligible firms |
| Handling in ChatGPT Enterprise | Use by eligible firms is excluded from human review by default |
Zero data retention in the API and exclusion from human review in ChatGPT Enterprise are different measures. The latter alone cannot be read as meaning that no data in ChatGPT is stored, and the same conditions do not necessarily apply to all users. OpenAI is working with Latham & Watkins on designs for access permissions and information barriers to prevent conflicts of interest. Client instructions and firm supervision are also within the scope of that effort.
ChatGPT for Word, which became generally available the same day, is a feature that supports document proofreading, suggested revisions, and flagging formatting issues. Its general availability should not be confused with Astra for Law itself having lost its usage restrictions.
The official help page asks users to "check the answer and its sources before relying on it." For those considering adoption in Japanese legal practice, in addition to confirming service regions and contract terms, it will be necessary to verify whether the tool can correctly read the grounds in the jurisdictions and matters actually handled. If those conditions are met, lawyers could build AI that supports everything from locating materials to applying them to a case into their own verification procedures.