OpenAI packages GPT-6 Astra for U.S. legal work

OpenAI introduced Astra for Law, a legal configuration of GPT-6 Astra that combines the model with a U.S. legal search index, OpenAI-maintained instructions and settings, firm workflow tools, and third-party integrations. The launch moves OpenAI deeper into the software stack used by large law firms, giving vendors and internal engineering teams a shared model and retrieval layer for legal products.

Where and when Astra ships

Initial access is limited to selected law firms in OpenAI’s Trusted Access program. API access is planned, although OpenAI has yet to disclose a release date, pricing, eligibility criteria, or usage limits.

The index carries the load

Legal research exposes a persistent weakness in general-purpose AI systems: a plausible answer can contain a nonexistent citation, miss controlling authority, or rely on a decision that an appellate court later reversed. Lawyers need the cited passage, the court and jurisdiction, publication status, and subsequent treatment of the decision.

Astra for Law indexes U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily. Through a partnership with the nonprofit Free Law Project, the index includes CourtListener’s collection, which the organization says covers more than 99.9% of published U.S. precedential case law.

OpenAI reports the following results on Vals AI’s Legal Research Bench:

The 54.0% correctness rate leaves 46.0% of benchmark questions failing the overall check. The results indicate stronger retrieval, but they do not support unsupervised use or replace a lawyer’s review of citations and subsequent case history.

OpenAI also presents a single head-to-head example in which Astra for Law found a relevant Southern District of New York precedent and a competing frontier model returned a decision that had been reversed on appeal. The example illustrates a concrete failure mode, although one comparison cannot establish how frequently either system makes that error.

Firms turn playbooks into software

Early law-firm collaborators have used OpenAI’s forward-deployed engineers to encode specific review standards, precedents, and drafting practices:

• Sullivan & Cromwell built an agreement analyzer that applies the firm’s negotiating playbooks and selected precedents, then produces proposed redlines and draft client advice.
• Ropes & Gray built a deal-diligence system that traces findings to source documents and identifies issues such as notice or consent requirements in customer contracts.
• Cooley built GO Public for IPO preparation, including updates that carry across a filing as the transaction changes.
• Wachtell, Lipton, Rosen & Katz is collaborating with OpenAI on tools intended to support complex legal analysis and judgment.
• Latham & Watkins is advising on governance, ethical walls, information permissions, and firm oversight.

These projects convert internal knowledge into repeatable workflows. The resulting tools can apply a firm’s review criteria consistently, preserve links to source material, and reduce the manual effort required to update drafts as a matter evolves.

OpenAI recruits the legal stack

The launch includes 26 partner-built plugins and nine community plugins covering 47 custom skills. Plugins connect ChatGPT to external systems, while custom skills package recurring tasks, instructions, and workflows for reuse.

• Thomson Reuters is bringing HighQ matter context into ChatGPT and previewing a CoCounsel Legal connector.
• iManage allows lawyers to save drafts to the relevant matter file.
• Relativity, Clio, Intapp, and DeepJudge provide connections to e-discovery, practice management, professional-services, and enterprise-search systems.
• Harvey and Legora plan to build on Astra for Law through the API.

Harvey has raised more than $1 billion while positioning its software as an AI layer for large law firms. Its participation shows how OpenAI can supply the underlying model and legal retrieval system to companies that already own the application interface, workflow design, and customer relationship. The same structure allows internal legal-engineering teams to build firm-specific tools against a maintained legal configuration.

Confidentiality controls set the boundary

For eligible firms, OpenAI says the API will support Zero Data Retention, a setting designed to prevent customer content from being stored after a request is processed. ChatGPT Enterprise usage is also excluded from human review by default.

Latham & Watkins is helping design controls for information permissions, ethical walls, client instructions, and firm oversight. Those controls matter in firms where lawyers working for different clients may need strict separation even when they use the same AI service.

Documents are still processed by OpenAI, and plugin calls may send content to connected vendors. Firms therefore need to assess privilege, contractual confidentiality, data residency, matter-level permissions, audit logging, client restrictions, and each integration’s retention policy. Zero Data Retention on the OpenAI API does not automatically govern data handled by a third-party plugin.

ChatGPT for Word became generally available alongside the launch, extending the product into the drafting environment. A compliant deployment depends on the firm’s architecture, vendor contracts, access controls, and review procedures across Word, ChatGPT, the API, and every connected service.

Who pays, who builds, who verifies

1. Legacy research vendors face a different buying equation. Westlaw and Lexis provide citators, editorial classification, secondary sources, and established research workflows. Astra’s URL count alone does not establish equivalent completeness or treatment analysis. Its broad index and daily updates could still strengthen firms’ leverage in procurement and reduce reliance on premium databases for some retrieval tasks.
2. Reasoning settings add cost and latency to billing decisions. Astra for Law is tuned for thorough work at high reasoning effort, which generally consumes more compute and takes longer to return an answer. Launch pricing remains undisclosed. Firms will need to measure whether research savings flow to clients through alternative fees or remain with the firm as margin.
3. The architecture can travel to other regulated fields. An OpenAI-maintained model configuration, specialized index, organization-specific workflows, integrations, and retention controls could support similar products in medicine, accounting, and finance.
4. Legal engineers gain a distribution channel. The community plugin program gives lawyers and developers a formal route for turning internal prompts, scripts, and review methods into reusable tools. Their work will include evaluation, permissions, source tracing, version control, and workflow design.

The API gaps that shape a build

For legal-technology developers, the planned API could reduce the work required to assemble a model, search index, legal instructions, and retrieval settings. OpenAI says it will maintain that configuration, allowing builders to concentrate on litigation, compliance, contract, and transaction workflows. The announcement leaves several implementation details unresolved:

• Access and economics: release date, pricing, rate limits, context window, throughput, and latency at each reasoning level.
• Retrieval output: stable source identifiers, quoted spans, court and jurisdiction metadata, publication status, and treatment such as reversal or limitation by a later court.
• Version control: options to pin model and index versions, receive change notices, test updates, and roll back regressions.
• Data rights: terms governing the display, storage, export, and reuse of retrieved legal text.
• Security: Zero Data Retention eligibility, data residency, audit logs, plugin permissions, and retention by connected services.
• Evaluation: performance on a developer’s own jurisdictions, document types, practice areas, and error thresholds.

Inside a firm, a practical pilot starts with a bounded workflow that has known source material and measurable outcomes. Suitable controls include matter-level access, source-linked outputs, mandatory citation verification, human approval before client delivery, and comparisons of accuracy, cost, and latency against the current process. The early projects from Cooley, Ropes & Gray, and Sullivan & Cromwell follow that pattern by applying encoded expertise to defined tasks while keeping lawyers responsible for the final work.