OpenAI is packaging its most capable model into a maintained legal system rather than asking every firm to assemble its own research stack from a general chatbot. GPT-6 Astra for Law combines the model with legal instructions, a large U.S. authority index, firm controls, and an ecosystem of specialist tools.
Watch the Official Astra for Law Video
Source and credit: Astra for Law: Frontier intelligence built for your practice, published by OpenAI on 17 September 2026. The accompanying official launch announcement is the source of record for product claims and availability.
What Astra for Law Actually Includes
Astra for Law is more than a system prompt and less than an autonomous lawyer. OpenAI describes it as a foundation for firms and legal-technology companies to build products and workflows around their own expertise.
| Layer | Role | What a firm must still own |
|---|---|---|
| GPT-6 Astra | Reasoning, synthesis, analysis, and drafting | Matter definition, supervision, and professional judgment |
| Legal instructions | Guide treatment of authority, counterarguments, uncertainty, and legal writing | Firm standards, jurisdiction-specific methods, and acceptance criteria |
| Legal Search Index | Find U.S. authorities and relevant passages | Shepardizing or equivalent treatment checks, binding status, and source verification |
| Firm context | Bring selected precedents, playbooks, documents, and matter data into workflows | Permissions, ethical walls, client instructions, retention, and data classification |
| Plugins and skills | Connect specialist products and repeatable legal workflows | Vendor review, account scope, action limits, and output review |
The Legal Search Index Is the Main Product Difference
The Legal Search Index covers U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs. OpenAI says new sources are added daily and that its work with Free Law Project brings CourtListener's collection, covering more than 99.9% of published U.S. precedential case law, into the research experience.
The index is designed to move from a factual pattern to candidate authorities and the passages that may matter. That is a better starting point than an unsupported answer, but the index does not remove the ordinary research questions: Is the authority binding? Is it still good law? Does it apply in the relevant jurisdiction? Is the quoted language part of the holding? Does a later source narrow the result?
What OpenAI's Benchmark Shows and What It Does Not
OpenAI evaluated the complete Astra for Law setup on 200 U.S. legal research questions from a private validation set of Vals AI's Legal Research Bench. At the highest reasoning effort, the company reports the following results:
| Reported measure | Astra for Law result | Comparison |
|---|---|---|
| Overall correctness check | 54.0% of questions passed | 38.7% for GPT-6 Astra with web search, a 40% relative improvement |
| Reference cases on case-law questions | 24% more cases found | Compared with GPT-6 Astra using web search at the same highest effort |
| Relevant passages from correct opinions | Up to 54% more | On the audited target passages at the same reasoning effort |
The test supports a narrow conclusion: OpenAI's legal configuration improved its own model's research performance on this private question set. It does not establish an error-free system, compare every specialist platform, measure every jurisdiction, or predict performance on a firm's confidential matters. A 54% pass rate also makes human verification an architectural requirement, not a disclaimer at the bottom of the page.
Where the System Could Fit in Legal Work
OpenAI's examples span litigation and transactional work. The model can search for factually similar authority, distinguish holdings from other observations, surface cases that weaken an argument, explain how a contract exception shifts risk, or turn research into a draft memo. Firms are also building narrower systems around their own standards:
- Agreement review: Sullivan & Cromwell built an analyzer that combines negotiating playbooks and selected precedents to find provision interactions, propose redlines, and draft advice for lawyer review.
- Deal diligence: Ropes & Gray built a system that traces data-room findings to source material and identifies issues such as notice or consent requirements.
- Capital markets: Cooley built GO Public to support IPO drafting, risk identification, and change propagation across a filing.
- Regulatory analysis: Skadden is developing tools to help clients assess regulatory risk around transactions and product launches.
These examples are firm-specific applications, not proof that one generic prompt can complete the same work. Their value comes from combining a capable model with selected sources, firm methods, permissions, interfaces, and a defined lawyer review process.
Access: Trusted Firms First, API Later
Astra for Law is initially offered to selected U.S. law firms through OpenAI's Trusted Access Program in ChatGPT and Codex. Access is for eligible lawyers and people working under their supervision. The model appears in the picker as GPT-6 Astra Law.
OpenAI says API access is coming soon and identifies the planned API model name as gpt-6-astra-law. That means builders should not design a production commitment around general API availability until access, pricing, limits, data handling, and regional coverage are documented for their account.
Legal-Grade Trust Requires More Than Model Quality
For eligible firms, OpenAI says the offering includes Zero Data Retention on the API, while ChatGPT Enterprise usage is excluded from human review by default. OpenAI is working with Latham & Watkins on information permissions, ethical walls, client instructions, and firm oversight.
| Control question | Required answer before a matter pilot |
|---|---|
| Who can see matter data? | Named users, groups, service accounts, vendors, and support paths |
| Where can the model search? | Approved public sources, licensed content, and specifically permitted firm repositories |
| What is retained? | Prompts, files, outputs, logs, embeddings, plugin data, and backups by product surface |
| What can plugins do? | Read, draft, save, modify, export, or transmit actions separated explicitly |
| How are ethical walls enforced? | Technical permissions tested against representative restricted matters |
| Who approves the output? | A qualified lawyer named in the workflow, with the source record preserved |
Zero Data Retention is important, but it is not the entire confidentiality model. Firms still need to understand connector behavior, enterprise logging, document-system retention, client consent or restrictions, cross-border requirements, and how administrators can audit use.
The Legal Plugin Ecosystem
OpenAI announced 26 partner-built legal plugins covering the practice and business of law. Examples include iManage for saving work to a matter file, Intapp for surfacing activity that may require a time entry, DeepJudge for comparing prior deals, Relativity and Clio integrations, and Thomson Reuters HighQ matter context. A CoCounsel Legal connector is described as forthcoming.
The community layer is easy to miscount. OpenAI says the launch includes nine community plugins with 47 custom skills from lawyers and legal engineers at LegalQuants, LECG, and Skills.law. The 47 figure is the number of adaptable skills, not 47 community plugins.
Harvey and Legora are named as API customers expected to build on Astra for Law. Their products remain separate specialist systems. The design principle is composability: let firms bring their tools and knowledge into a governed environment rather than forcing every workflow into one vendor's interface.
A Lawyer Review Loop for Every Research Answer
- Freeze the question. Record the client facts, jurisdiction, date, research scope, and excluded issues.
- Require authority-level citations. Ask for the court, date, reporter or docket, relevant passage, and an explanation of why it matters.
- Open every material source. Confirm that the cited passage exists and supports the proposition attributed to it.
- Check treatment and hierarchy. Verify good-law status, binding weight, later authority, and jurisdictional fit in an approved system.
- Search against the answer. Ask for adverse authority, contrary interpretations, missing elements, and facts that would change the conclusion.
- Separate model text from lawyer judgment. Keep the generated draft, revisions, source record, and final approved work product distinguishable.
- Protect the matter. Confirm client instructions, privilege, ethical walls, connector permissions, and retention before adding confidential content.
A Safe 30-Day Firm Pilot
Start with a bounded research or drafting workflow where errors are visible and no external action occurs automatically. A useful pilot could compare Astra for Law with the firm's current process on 20 completed matters whose authoritative answer and research trail are already known.
- Measure accepted accuracy: correct authorities, passages, treatment, and factual application after lawyer review.
- Measure completeness: important adverse authority, missing issues, and questions the model should have asked.
- Measure effort: research time, source-check time, corrections, and senior-review time.
- Test controls: restricted matters, blocked repositories, plugin scopes, audit logs, and attempted cross-matter retrieval.
- Set a stop rule: pause the pilot if confidentiality boundaries fail, citations cannot be reproduced, or review effort exceeds the baseline.
The adoption decision should be based on accepted work per lawyer hour, not how polished the first draft looks.
Official Sources and Useful Links
- OpenAI: Astra for Law official video
- OpenAI: Introducing Astra for Law
- OpenAI Help: Astra for Law access, model name, review guidance, and plugins
- OpenAI: solutions for law firms
- OpenAI: legal plugins for ChatGPT and Codex
The official video and announcement were published on 17 September 2026. This article was reviewed on 23 September 2026. Access, plugin availability, API timing, pricing, benchmark configurations, and safeguards may change.