OpenAI Attached a 230-Million-URL US Legal Index to GPT-6 Astra and Called It Astra for Law

OpenAI introduced Astra for Law on September 17, 2026, combining GPT-6 Astra with a legal search index and custom instructions for legal analysis and writing. The index covers US case law, statutes, regulations, court rules and administrative decisions across a corpus of more than 230 million URLs with sources added daily, and a collaboration with Free Law Project, the nonprofit behind CourtListener, brought in a case-law collection covering more than 99.9% of published US precedential case law. On 200 US legal research questions from the private validation set of Vals AI's Legal Research Bench, Astra for Law passed the overall correctness check on 54.0% of questions against 38.7% for GPT-6 Astra using web search alone, a 40% relative improvement. API customers including Harvey and Legora will build on the configuration, and 26 partner-built plugins shipped alongside it. ASAP separates out why the index, not the model, is the subject of this announcement.

Astra for Law is a configuration of GPT-6 Astra, not a separately trained legal model

Astra for Law is not a separately trained legal model but GPT-6 Astra with a legal search index and legal analysis and writing instructions layered on top, surfacing as "GPT-6 Astra Law" in the model picker and as gpt-6-astra-law in the API. OpenAI describes it as its most powerful model configured into a new AI foundation for law. There are three pieces: the frontier model GPT-6 Astra, a search index over case law and statutes, and instructions that specify how legal analysis and writing should be done.

The structure carries an explicit succession plan. OpenAI states that as its frontier models advance, it will bring these legal capabilities to its latest models. Because the legal capability lives outside the weights, in the index and the instructions, swapping the underlying model leaves the legal configuration intact.

Availability is still narrow. Astra for Law is initially offered to selected law firms through Trusted Access in ChatGPT and Codex, with the API described as coming soon. The same day, OpenAI made ChatGPT for Word generally available so lawyers can proofread, get suggested edits and flag formatting issues inside Word.

The 230-million-URL index and CourtListener are the real assets in this release

The legal search index spans US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with sources added daily. OpenAI states explicitly that the index complements, rather than replaces, the licensed content and specialist products firms rely on from providers such as Thomson Reuters.

The case-law portion comes from Free Law Project, the nonprofit behind CourtListener. OpenAI says the collaboration brings that organization's case-law collection, covering more than 99.9% of published US precedential case law, into the research experience.

The workflow OpenAI describes explains why the asset comes first. Legal research typically begins with finding the exact right authority, locating the relevant passages, and understanding how relevant and binding they are to the situation at hand. The index is the tool for that work, and the model is the thing that uses it.

54.0% against 38.7% compares two settings of the same model, not two companies

The 54.0% and 38.7% figures OpenAI published do not pit one company's model against another's; they compare the same GPT-6 Astra running in the legal configuration against the same model running with web search alone. The evaluation used 200 US legal research questions drawn from the private validation set of Vals AI's Legal Research Bench, with both systems set to the highest reasoning effort. The benchmark measures how well the model finds relevant sources and passages and how well its research answers meet the evaluation criteria.

Retrieval numbers were published separately for case-law-focused questions. At the highest reasoning effort, Astra for Law found 24% more reference cases than GPT-6 Astra using web search alone, and on the audited set of target passages it retrieved up to 54% more relevant passages from the correct court opinions when the systems were compared at the same reasoning effort.

One more reading of the numbers keeps the picture accurate. The 54.0% figure is the endpoint of a 40% relative gain, but it also means nearly half the questions still fail the overall correctness check. And the qualifier "up to" on passage retrieval points at a ceiling, not an average. Taken together, the result reads less as a declaration that legal research is solved and more as an observation that attaching a corpus to an unchanged model moved the number by 15.3 percentage points.

By naming Harvey and Legora as customers, OpenAI moved onto the application layer's floor

The sentence naming Harvey and Legora as API customers sits where a declaration of non-competition with the legal AI application layer would go, and OpenAI writes that building on OpenAI should mean getting more from the ecosystem, not replacing it. The plugin list points the same direction. The 26 partner-built plugins connect ChatGPT to tools such as iManage, Intapp and DeepJudge, and Thomson Reuters is bringing HighQ matter context into ChatGPT while previewing a forthcoming CoCounsel Legal connector.

Facts pointing the other way sit in the same announcement. OpenAI's forward-deployed engineers adapted ChatGPT Enterprise with custom interfaces for individual firms, producing Sullivan & Cromwell's agreement analyzer, Ropes & Gray's deal diligence system and Cooley's GO Public. All three are the kind of workflow product legal AI companies have been selling.

Overlay the two and the position of this launch becomes visible. OpenAI is selling the lower layer of model and index while simultaneously building the application layer directly with selected firms. Thomson Reuters CTO Joel Hron's quote, which opens by saying the value is not in connectivity alone and closes by asserting that CoCounsel remains the trusted professional AI system designed to help complete that work, reads as a partner redrawing the boundary of its own territory. For Harvey and Legora, this launch is a better engine and, at the same time, an event that shifts their accumulated differentiation toward the index and the instructions.

The first thing a firm checks is not the pass rate but the data-handling terms

What actually decides adoption here is not the 54.0% pass rate but the data-handling terms attached to the Trusted Access Program, which for eligible firms includes Zero Data Retention on the API and ChatGPT Enterprise usage excluded from human review by default. OpenAI describes the program as a separate path for lawyers and people working under their supervision to use the model for professional legal work.

Read it alongside the four items OpenAI says it is designing with Latham & Watkins: information permissions, ethical walls, client instructions and firm oversight. That this list contains no performance items at all, only access-control items, reveals where the bottleneck in this product sits. Which lawyer inside a firm can reach which matter's material is not a convenience question but a precondition of doing the work, and a model with access to an organization's entire document store runs straight into that precondition.

The evaluation order therefore differs from ordinary enterprise adoption. Reading the benchmark first and checking the controls later does not produce a decision. Whether the control terms satisfy a given firm's requirements is settled first, and only then do retrieval gains of 24% and 54% carry meaning.

The index covers US law only, which sets the terms for other jurisdictions

Astra for Law's search index covers US case law, statutes, regulations, court rules and administrative decisions, and none of that transfers directly to legal practice in Korea or any other jurisdiction. If the measured gain came from the corpus and instructions rather than the weights, then what is reproducible elsewhere is the method, not the product.

The transferable part is fairly clear: assemble a verified corpus of statutes and case law, attach a retrieval tool over it, and write separate drafting instructions covering things like distinguishing a holding from a court's other observations or addressing cases that weaken an argument. The instruction examples OpenAI cites live at exactly this layer, including distinguishing a court's holding from its other observations, addressing cases that weaken an argument, and explaining how a contract exception shifts risk between the parties.

The part that does not transfer is the corpus. OpenAI reached its 99.9% coverage through a collaboration with CourtListener, an open nonprofit project, not through its own collection effort. The first question for a team building the same configuration in another jurisdiction is not which model to use but who opens an equivalent corpus and under what terms. Without an answer there, no amount of model swapping produces the 15.3 percentage points this launch demonstrated.

The competitor comparison is two examples, and the benchmark set is private

The source for these numbers is the private validation set of Vals AI's Legal Research Bench, and no third-party reproduction under the same conditions accompanied the announcement. A private validation set guards against benchmark contamination, but it also means readers cannot check the same questions themselves.

The competitor comparison is a different kind of evidence. OpenAI showed side-by-side answers from Astra for Law and Claude Fable 5.1 on one litigation prompt and one transactional prompt, reporting that Fable 5.1 returned a holding that had been reversed on appeal in the litigation example and reported finding no such case in the transactional example. These are two examples selected by the announcing company, not an aggregated evaluation. The format leaves room to pick favorable cases, so it is weak ground for a judgment about the competitive picture.

Several items remain unconfirmed. First, neither the number of firms receiving Trusted Access nor the selection criteria were disclosed, and the API timeline is given only as "coming soon." Second, the 9 community plugins and 47 custom skills are described as the work of lawyers and legal engineers at LegalQuants, LECG and Skills.law, with no quality-verification process explained. Third, the index is described as updated daily, but neither the update lag nor the method for checking whether a case is still good law is specified. Given that determining whether a precedent remains good law is the core risk in this work, the third item is the question practitioners should ask first.

Source: Introducing Astra for Law (OpenAI, September 17, 2026)