Astra for Law: OpenAI Turns GPT-6 Astra Into a Legal Research Machine With a 230M-URL Index

On September 17, 2026, OpenAI made its most direct move yet into one of the last high-margin professional markets: it launched Astra for Law, a legal configuration of its GPT-6 Astra flagship model, complete with a dedicated legal search index covering more than 230 million URLs of United States case law, statutes, regulations, court rules, and administrative decisions.

The company called it “the beginning of its long-term investment in law,” and the shape of the announcement tells you why. This is not a fine-tuned model or a standalone product. It is a vertical platform play: a model configuration, a retrieval layer, a plugin ecosystem, and an enterprise access program bundled into one offering aimed simultaneously at large law firms and the legal technology vendors that serve them.

What Astra for Law Actually Is

At a media briefing, OpenAI’s Jason Boehmig — the Ironclad cofounder hired to lead the company’s legal products — emphasized that Astra for Law is not a new model. It is GPT-6 Astra with domain-specific instructions for legal writing and analysis, response-length settings tuned for professional work, and tooling specific to the legal industry.

The centerpiece of that tooling is the Legal Search Index. Built in partnership with the nonprofit Free Law Project and drawing on CourtListener’s collection — which its maintainers say covers more than 99.9% of published U.S. precedential case law — the index is updated daily with new sources. OpenAI says the system is designed to go beyond retrieving law: it should help lawyers apply the law to the facts of a matter, develop arguments or deal terms, and identify weaknesses and uncertainty in their own positions.

At the briefing, OpenAI’s Ansel Wu demonstrated three tasks: drafting a motion to dismiss for a hospital defending a disability-discrimination and retaliation claim, a stockholder challenge to an asset sale rushed through over a holiday weekend, and a sales-commitment dispute.

The Numbers: 54% vs 38.7%

OpenAI tested the system against 200 U.S. legal research questions from the private validation set of the Vals AI Legal Research benchmark, which measures whether a model can find relevant sources and passages and whether its answers meet evaluation criteria.

At the highest reasoning effort:

• Astra for Law passed the overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone — a 15.3 percentage-point gain, or roughly a 40% relative improvement.
• On case-law-focused questions, it found 24% more reference cases than the web-search baseline.
• On audited target passages, it retrieved up to 54% more relevant passages from the correct court opinions.
• Its answers were roughly twice as long on average as the baseline across tested reasoning settings.

The 46% failure rate deserves attention too. As AlphaSignal’s analysis noted, stronger retrieval does not support unsupervised use, and it does not replace a lawyer’s review of citations and subsequent case history. OpenAI itself presented a head-to-head example in which Astra for Law found a relevant Southern District of New York precedent while a competing frontier model returned a decision that had been reversed on appeal — a single anecdote, but a vivid illustration of why retrieval grounding matters in law.

Plugins, Skills, and the Ecosystem Play

The launch shipped with an entire ecosystem attached:

• 26 partner-built plugins from Thomson Reuters, Harvey, Legora, iManage, Relativity, Clio, Intapp, DeepJudge, and others. iManage’s plugin saves drafts to the matter file; Intapp’s surfaces activities needing time entries; Thomson Reuters’ brings HighQ matter context into ChatGPT, with a CoCounsel Legal connector to follow.
• 9 community plugins built by lawyers and legal engineers, plus 47 custom skills developed by legal “power users,” not by OpenAI.

Alongside the launch, OpenAI announced general availability of ChatGPT for Word, letting lawyers proofread, get suggested edits, and flag formatting issues inside the drafting environment where legal work actually happens.

The firm-level stories are the most telling. Sullivan & Cromwell built an agreement analyzer that applies its negotiating playbooks and selected precedents to produce redlines and draft client advice. Ropes & Gray built an M&A diligence system that traces findings back to source documents in the data room. Cooley built “GO Public,” encoding its capital-markets expertise into IPO preparation from drafting the filing to surfacing the risks that deserve management attention. All were built with OpenAI’s forward-deployed engineers embedded in the firms.

Trusted Access: Zero Data Retention for Am Law 200

Initial access runs through a new Trusted Access program targeting Am Law 200 firms, granting specialized access through ChatGPT and Codex. For qualifying firms, the program includes zero data retention on the API and exclusion of ChatGPT Enterprise usage from human review by anyone at OpenAI.

Latham & Watkins is working with OpenAI to design information permissions, ethical walls, client instructions, and firm oversight — controls that matter in firms where lawyers for different clients need strict separation even on shared infrastructure. When a reporter asked about exceptions to the zero-retention policy, Boehmig acknowledged the one-sentence summary understates a roughly 30-page access agreement, but said OpenAI is “extremely confident” the terms meet the industry’s needs.

API access for vendors including Harvey and Legora is coming soon, though OpenAI has not disclosed release dates, pricing, eligibility criteria, or usage limits.

Analysis: Coopetition With Harvey and the Vertical Stacking of AI

The most strategically loaded fact in the announcement is who gets API access: Harvey, the legal AI company that has raised more than $1 billion building an application layer on top of OpenAI’s models, and Legora. OpenAI is now supplying the model, the legal retrieval index, and the instructions to the very companies whose differentiation was assembling those pieces themselves.

The likely equilibrium is coopetition: Harvey keeps the interface, the workflow design, and the customer relationship, while OpenAI maintains the substrate. But every layer OpenAI absorbs raises the same question that software platforms have asked their ecosystem for decades — what happens to the margin of the layer above?

The competitive threat to legacy research vendors Westlaw and Lexis is more nuanced. Their citators, editorial classification, and secondary sources are not replicated by a URL count, and Astra’s index breadth alone does not establish equivalent treatment analysis. But a maintained, daily-updated legal index available at API prices strengthens every firm’s negotiating position in procurement, and could pull commodity retrieval work away from premium databases.

There is also a timing dimension: Anthropic announced its own major legal push earlier, with Claude for Legal targeting the same buyers. Novo Nordisk’s Claude adoption for drug discovery the same week shows the vertical playbook spreading across regulated industries. The architecture here — model configuration, specialized index, firm-built workflows, integrations, retention controls — is portable to medicine, accounting, and finance.

The Caveats That Matter

AlphaSignal’s launch analysis flagged the unresolved details worth watching: no disclosed release date, pricing, rate limits, context window, or latency figures; no stable source identifiers or pinned model/index versions for reproducible citations; terms for display and reuse of retrieved legal text; and the fact that zero data retention on the OpenAI API does not govern what third-party plugins do with content sent to connected vendors.

For a profession where a fabricated citation can mean sanctions, the burden of proof remains on verification. A 54% correctness rate is a major step up from 38.7% — and a long way from a system lawyers can trust without checking.

But the direction is unmistakable. The general-purpose chatbot era of legal AI is ending; the vertically integrated platform era has begun. OpenAI has just made clear it intends to own a large share of the legal stack — and it is starting at the top of the market, where the tolerance for cost is highest and the tolerance for error is lowest.