OpenAI Astra for Law: Legal Research, Workflows, and Access

OpenAI Astra for Law is a legal AI foundation that combines GPT‑6 Astra with legal search, instructions for legal analysis and writing, governance controls, and integrations with specialist legal tools. It is intended for law firms and legal technology companies building research, drafting, diligence, and knowledge workflows around their own expertise.

OpenAI announced the product on September 17, 2026. Initial access will be limited to selected eligible law firms through a Trusted Access Program in ChatGPT and Codex. OpenAI says API access is coming soon under the model name gpt-6-astra-law; the ChatGPT and Codex model picker will show “GPT‑6 Astra Law.” The announcement does not provide a general-availability date or public pricing.

Source note: Product details and benchmark figures in this article come from OpenAI’s official Astra for Law announcement. The reported evaluation uses a private validation set and was published by OpenAI. Results on a firm’s matters, jurisdictions, documents, and approved research sources may differ. This article provides technical and operational analysis, not legal advice.

Astra for Law at a glance

What is Astra for Law?

Astra for Law is a configuration of OpenAI’s frontier model and legal tooling for professional legal work. Its three central components are GPT‑6 Astra, a legal search index, and custom instructions that guide legal research, analysis, and writing.

That distinction matters. A general model with web search can find public pages, but legal work requires more than topical relevance. A lawyer must identify controlling and persuasive authority, inspect the exact passage, determine whether the authority remains good law, understand the procedural and jurisdictional context, and apply it to the client’s facts.

OpenAI positions Astra for Law as a foundation that firms and legal technology companies can build around. API customers including Harvey and Legora are named in the announcement, while firms can connect their own knowledge, tools, permissions, and review processes.

How Astra for Law’s legal search works

OpenAI says the legal search index covers U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with new sources added daily.

The company also says its work with the nonprofit Free Law Project brings CourtListener’s collection into the research experience. OpenAI cites Free Law Project’s estimate that the collection covers more than 99.9% of published U.S. precedential case law.

The index is not presented as a substitute for every licensed source. OpenAI specifically says it complements licensed content and specialist products from providers such as Thomson Reuters. A firm should define which sources are permitted for each workflow and when a result must be checked in its authoritative research platform.

Search coverage is not the same as legal validity

A large corpus can improve retrieval without proving that an answer is legally sound. Production workflows still need checks for:

• jurisdiction and court hierarchy;
• publication and precedential status;
• subsequent history and negative treatment;
• effective dates and amendments;
• quotations and pinpoint citations;
• factual similarity and material distinctions;
• privileged, confidential, sealed, or restricted information.

The safest interface makes the source and cited passage easy for a lawyer to inspect. It should never hide uncertainty behind a polished memorandum.

What OpenAI’s benchmark results show

OpenAI tested the complete Astra for Law configuration on 200 U.S. legal research questions from a private validation set of Vals AI’s Legal Research Bench. At the highest reasoning effort, OpenAI reports that Astra for Law passed the evaluation’s overall correctness check on 54.0% of questions, compared with 38.7% for GPT‑6 Astra using web search alone.

That is a 15.3 percentage-point difference and a 40% relative improvement. OpenAI also reports that, on case-law-focused questions at the highest reasoning effort, Astra for Law found 24% more reference cases. On an audited set of target passages, it retrieved up to 54% more relevant passages from the correct court opinions when the two systems used the same reasoning effort.

These results support the value of legal-specific retrieval and instructions. They do not establish error-free research. A 54.0% pass rate on the reported correctness check also means the tested configuration did not pass that check on 46.0% of questions. The private validation set limits independent inspection, and U.S. benchmark performance cannot be assumed to transfer to Canadian law, other jurisdictions, firm-specific documents, or every practice area.

Beyond research: analysis, drafting, and deal workflows

OpenAI says Astra for Law includes custom instructions for applying research to client facts, developing arguments or deal terms, distinguishing holdings from other observations, addressing adverse authority, and explaining how contract language changes risk.

The announcement describes several examples developed with law firms:

• Sullivan & Cromwell built an agreement analyzer that applies negotiating playbooks and selected precedents, identifies interactions between provisions, and prepares proposed redlines and draft client advice for lawyer review.
• Ropes & Gray built a deal-diligence system that follows findings back to source documents and identifies questions that may affect an acquisition.
• Cooley built GO Public to support IPO preparation, including drafting, risk identification, and carrying deal changes across a filing for review.

These examples show the broader opportunity: a useful legal AI system combines the model with the firm’s own standards, approved knowledge, matter context, permissions, and review steps. The value comes from the complete controlled workflow, not the model name alone.

Trusted Access, confidentiality, and governance

OpenAI created a Trusted Access Program for eligible law firms and people working under lawyer supervision. The company says the offering includes Zero Data Retention on its API for eligible firms, while usage of ChatGPT Enterprise is excluded from human review by default.

Those controls are important, but a firm still needs its own legal, security, privacy, records, and professional-responsibility review. Before placing client information into any AI workflow, confirm the exact contract and configuration rather than relying on a product-page summary.

A production approval should cover:

1. Matter-level access: Who can see each client, matter, document, and generated output?
2. Ethical walls: Can conflicting teams or users retrieve one another’s information?
3. Client instructions: Has the firm respected client-specific restrictions on AI, hosting, and third parties?
4. Data handling: What is retained, logged, reviewed, cached, backed up, or sent through plugins?
5. Source policy: Which public, licensed, and internal sources are approved for each task?
6. Human responsibility: Which lawyer verifies and approves the work before it affects advice, filings, negotiations, or clients?
7. Evidence and audit: Can the firm reproduce the prompt, sources, tool calls, model version, output, edits, and final approval?

OpenAI says it is working with Latham & Watkins on information permissions, ethical walls, client instructions, and firm oversight. Firms should still validate the controls in their own environment and professional context.

Plugins, custom skills, and ChatGPT for Word

OpenAI is launching 26 partner-built legal plugins that connect ChatGPT to specialist legal and business tools. The announcement names examples involving Relativity, Clio, iManage, Intapp, DeepJudge, Thomson Reuters HighQ, and a forthcoming CoCounsel Legal connector.

It also includes nine community plugins created by lawyers and legal engineers, with 47 custom skills that practitioners can adapt or use as examples. Separately, OpenAI says ChatGPT for Word is generally available, offering proofreading, suggested edits, and formatting checks inside a drafting environment lawyers already use.

Connections create practical value, but they also extend the security boundary. Every plugin should be evaluated for identity, permissions, tenant isolation, source access, write actions, data retention, logs, failure behaviour, and revocation. A convenient connection should not become unrestricted access to a document management system, time records, discovery platform, or client matter.

Astra for Law versus general GPT‑6 Astra

The core difference is the legal system around the model. GPT‑6 Astra is the underlying frontier model. Astra for Law adds legal search, legal-specific instructions, access controls, and ecosystem integrations intended for professional workflows.

A firm should choose based on the workflow, evidence requirements, and approved information environment—not assume that a specialized label removes the need for evaluation.

A safer implementation plan for law firms

1. Start with one narrow, reviewable workflow

Choose a task with clear inputs and outputs, such as identifying candidate authorities, comparing defined clauses, preparing a first-pass chronology, or flagging formatting problems. Do not begin with autonomous legal advice or uncontrolled access to every matter.

2. Build a representative private evaluation

Use completed matters, synthetic examples, known difficult cases, adverse authorities, superseded law, ambiguous clauses, incomplete records, and jurisdictional edge cases. Measure citation accuracy, passage accuracy, issue coverage, hallucinations, lawyer correction time, and failure severity.

3. Separate research from final judgment

Let the system retrieve, organize, compare, and draft. Require a qualified professional to inspect the authorities, factual assumptions, analysis, and final work product.

4. Apply least-privilege access

Connect only the matters, sources, plugins, and write actions needed for the pilot. Use separate identities and environments for testing and production. Keep sending, filing, publishing, redlining shared documents, and changing records behind explicit approval.

5. Verify the downstream result

A completed model response is not proof that a document was saved correctly, a redline contains the approved changes, a citation is current, or a filing reached the right destination. Check the final state in the authoritative system.

6. Monitor versions and drift

Record the model, instructions, sources, plugins, and evaluation version. Re-test when OpenAI updates the model, index, legal instructions, integrations, or access terms.

Important limitations

Astra for Law is a newly announced product with limited initial access. The public evidence currently comes mainly from OpenAI’s own announcement, partner comments, and a private benchmark validation set. General availability, public pricing, final API documentation, rate limits, and complete regional or jurisdictional coverage were not published in the announcement.

The legal search coverage described by OpenAI is U.S.-focused. Organizations working in Canada or other jurisdictions should not infer equivalent coverage. Every firm must also consider professional conduct rules, privilege, confidentiality, court requirements, client terms, vendor management, and the laws that apply to its practice and data.

How I can help build a controlled legal AI workflow

I provide AI consulting and custom development for organizations evaluating agent systems, private knowledge, document workflows, permissions, human review, and production validation.

I can build the surrounding product through workflow automation, SaaS product engineering, website development, and mobile app development. The XReporter operations and reporting system demonstrates how structured information, controlled workflows, and clear reporting can work together.

Book a free strategy call to define one narrow legal or professional-services workflow, its access boundaries, and the evidence required before automation is trusted.

Official OpenAI sources

• OpenAI: Introducing Astra for Law
• OpenAI solutions for law
• OpenAI legal plugins