OpenAI launches Astra for Law with a 230-million-page U.S. legal index

Introducing Astra for Law: Scope and Architecture

OpenAI has unveiled a new product called Astra for Law, which integrates the GPT-6 Astra model with a specialized index and a set of instructions that are tailored to the requirements of professional legal work in the United States. The announcement positions the platform as a dedicated environment for legal research and drafting, distinct from the broader ChatGPT offering.

The core of the platform’s knowledge base is an index that spans more than 230 million pages of American legal material. This material includes case law, statutes, regulations, procedural rules, and administrative decisions, all of which are relevant to the practice of law in the United States.

OpenAI’s partnership with the Free Law Project extends the coverage of the index to an announced 99.9 % of published American precedent. The collaboration is presented as a way to ensure that the vast majority of binding case law is searchable within the system.

Performance Benchmarks and Comparative Results

In an internal evaluation covering 200 questions of American law, OpenAI reports an overall success rate of 54 % for Astra for Law. By contrast, a version of GPT-6 Astra that relied only on a simple web search achieved a success rate of 38.7 %. These figures are supplied by OpenAI and are not independently verified.

When the test set is limited to questions that focus on legal precedents, OpenAI states that Astra for Law provides 24 % more relevant case citations than the baseline web‑search approach. The same subset also shows up to a 54 % improvement in the system’s ability to locate pertinent passages within those citations.

The reported gains suggest that the combination of a dense legal index and model‑level instructions can enhance both the breadth of retrieved authorities and the precision with which specific passages are identified. However, the methodology behind the success metrics is not disclosed, leaving open questions about the difficulty of the questions, the scoring rubric, and the potential influence of prompt engineering.

Adoption by Law Firms and Integration Ecosystem

Sullivan & Cromwell has built a contract‑review tool that incorporates its own guidelines and precedent collections into Astra for Law. The firm’s involvement illustrates one concrete use case where a large practice integrates the platform with internal standards.

Other law firms are reported to be developing applications for due‑diligence processes and for supporting initial public offerings. These initiatives indicate a broader interest in leveraging the platform for tasks that require systematic analysis of large bodies of legal text.

Astra for Law also includes a suite of 26 partner plugins that connect ChatGPT to external legal‑technology tools. The presence of these plugins suggests that the platform can be extended to workflows such as document management, citation checking, and case‑law analytics.

For clients that access the service through the API, OpenAI offers an option for zero data retention. This feature is intended to address confidentiality concerns that are central to legal practice, although the exact technical guarantees are not detailed.

Access Model and Professional Responsibility

Currently, access to Astra for Law is limited to a select group of law firms via the ChatGPT and Codex interfaces. A broader API rollout is planned, but the timeline and criteria for eligibility have not been disclosed.

OpenAI explicitly states that the tool does not eliminate the lawyer’s duty to verify the output. The platform is positioned as an assistive technology rather than a substitute for professional judgment, a stance that aligns with existing ethical guidelines for legal practice.

• Integration of a proprietary legal index with GPT‑6 Astra
• Partnership with the Free Law Project for near‑complete precedent coverage
• Reported performance improvement over a simple web‑search baseline
• In‑house tools built by firms such as Sullivan & Cromwell
• Zero‑retention option for API customers

The limited release model raises questions about the scalability of the platform’s infrastructure, particularly given the size of the underlying index. It also prompts inquiry into how the system will handle updates to statutes and new case law as they are published.

Future research could examine whether the observed gains in citation relevance and passage retrieval persist across different legal domains, such as tax law or intellectual property, where the structure of authorities differs from general case law.

If the platform’s capabilities continue to improve, it may influence the cost structure of legal research services, potentially reducing the time lawyers spend on preliminary document review. Such a shift could, in turn, affect billing models and the allocation of junior staff to research tasks.

Nevertheless, the reliance on a proprietary AI system introduces new risk vectors, including model bias, data security, and the need for ongoing validation of outputs. These considerations suggest that the legal profession will need to develop new oversight mechanisms to manage AI‑assisted work.