Astra for Law: Why Legal AI Is Becoming a Working System

A lawyer is staring at a half-finished memo at 6:40 p.m. The client does not need a paragraph that merely sounds confident. The lawyer needs the right case from the right court, the passage that actually matters, an explanation of whether that authority controls, and a memo that admits where the argument is vulnerable.

That gap is the story behind Astra for Law, introduced by OpenAI on September 17, 2026. It combines GPT-6 Astra, a high-capability general-purpose model, with a legal search index, instructions for legal analysis and writing, privacy controls, and connections to the tools law firms already use. The important idea is not a chatbot with a legal name. It is a legal work system built around research, context, and review. (openai.com)

A model is only one piece

A foundation model is an AI model trained to handle many kinds of language and reasoning tasks. It can summarize a contract, explain a doctrine, or draft a letter. But professional legal work adds requirements that ordinary conversation does not: jurisdiction, procedural posture, authority, confidentiality, and a clear trail back to source material.

Astra for Law treats the model as one part of a larger setup. A useful way to picture the workflow is:

client facts
 ↓
legal search index
 ↓
authorities and relevant passages
 ↓
legal analysis instructions
 ↓
draft memo with weaknesses and uncertainty
 ↓
lawyer review

That structure matters because finding a plausible answer is different from supporting a legal position. The system is designed to help connect a client’s facts to authorities, then explain how closely those authorities fit.

Search first, then reason

Legal research depends on more than matching words. An authority is a source recognized by law, such as a statute, regulation, court opinion, or administrative decision. A court’s holding is the rule necessary to decide the case; other observations may be useful but not binding. A strong research workflow needs to distinguish those pieces instead of treating every search result as equally important.

The practical search question is: how does Astra for Law improve legal research beyond a general chatbot with web access? Its legal search index is built to search U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with new sources added daily. OpenAI also says its work with Free Law Project brings CourtListener’s collection, covering more than 99.9% of published U.S. precedential case law, into the research experience. (openai.com)

That does not replace licensed legal databases or a firm’s existing research products. It gives the model a more appropriate starting point: a large legal corpus, targeted passages, and instructions that emphasize authority and factual fit. The lawyer still needs to open the opinion, check the current status of the law, and decide whether a case truly belongs in the memo.

What the early benchmark numbers mean

OpenAI evaluated 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 setting, Astra for Law passed the benchmark’s overall correctness check on 54.0% of questions, compared with 38.7% for GPT-6 Astra using web search alone. OpenAI describes that as a 40% relative improvement, which is also a 15.3 percentage-point increase.

The system also found 24% more reference cases on case-law-focused questions and retrieved up to 54% more relevant passages from the correct opinions in an audited set. Those results suggest that legal retrieval and source selection can materially change the quality of an answer. They do not mean that 54% of legal answers are safe to file without review, nor do they turn a benchmark into a guarantee of legal correctness. A benchmark measures performance under a defined test; a real matter contains messy facts, incomplete records, and consequences that do not fit neatly into a score.

From a search result to a legal argument

The launch gives a useful example involving a manufacturing contract. A client was told that comparable suppliers historically received about $8 million in annual orders, but the signed agreement guaranteed only $2 million and left additional purchases to the retailer’s discretion. The retailer was also changing its allocation process.

A shallow answer might treat the historical figure as a promise of future business. Astra for Law instead separated two ideas: whether the retailer owed $8 million in orders, and whether a false statement about past order volume helped induce the client to sign. That second theory is a potential misrepresentation claim, meaning an allegation that a material false statement caused someone to enter a transaction.

The example also shows why factual similarity matters. Astra for Law identified a closer precedent involving representations about historical compensation during a changing business system, then pointed out the weak point: the client already knew the allocation process was changing. That fact could undermine reasonable reliance, the requirement that the client’s decision to trust the statement was legally justified. The result is more useful than a case name alone because it explains both the match and the fracture in the analogy.

Governance belongs inside the workflow

Legal AI cannot be judged only by how persuasive its prose sounds. Client files may contain privileged communications, trade secrets, personal data, and information that must not cross between matters. An ethical wall is a set of permissions and procedures designed to prevent confidential information from one client or team from reaching another.

For eligible firms, OpenAI’s current Astra for Law program includes Trusted Access, Zero Data Retention on the API, and default exclusion of ChatGPT Enterprise usage from human review. OpenAI also says it is working with Latham & Watkins on information permissions, ethical walls, client instructions, and firm oversight. These controls are not administrative decoration. An AI system that finds the right case but exposes the wrong client file has still failed at legal work.

The same principle appears in the tool connections around Astra for Law. OpenAI says the launch includes 26 partner-built plugins, with examples including iManage, Intapp, DeepJudge, and Thomson Reuters HighQ, along with community-built skills from legal professionals and engineers. The aim is a composable setup: firms can connect approved knowledge and matter systems without pretending that one model should replace every specialist product.

The final handoff remains human

Astra for Law can improve the first pass through research and drafting, but the final responsibility still belongs with the legal team. Before relying on an answer, a lawyer needs to verify the cited authority, read the relevant opinion, confirm that the rule is current and binding in the jurisdiction, test the factual analogy, and check that the data used by the system was permitted for that matter. OpenAI’s own Help Center tells users to review answers and cited sources before relying on them. (help.openai.com)

As of the current rollout, Astra for Law is being offered first to selected U.S. law firms through Trusted Access in ChatGPT and Codex, with API availability described as coming soon. That limited release fits the product’s larger message: legal AI is moving toward carefully governed systems that combine model capability with evidence, firm expertise, and accountable review.

The most meaningful shift is conceptual. Better legal AI is not only about producing smoother sentences. It is about helping a lawyer move from client facts to supported authority, from authority to analysis, and from analysis to a decision that a human professional can inspect and defend.