Astra for Law Launches: OpenAI's Legal Tool Adds an Index, Not a New Model

On September 17, 2026, OpenAI published the announcement “Introducing Astra for Law.” OpenAI's own framing is that it configured “our most powerful model” into a new AI foundation for law: combining GPT‑6 Astra with settings, tools, and context built for professional legal work. The same day, OpenAI also published a customer story about the international law firm Cooley, which built GO Public using a different product, ChatGPT Work.

This article was checked on September 18, 2026, against OpenAI's official announcement page, the Cooley customer story page, and the legal category of OpenAI's plugin directory. We have not used Astra for Law ourselves; the feature descriptions, benchmark figures, and customer statements below are all OpenAI's own claims or those of its customers. This article is not legal advice, and we do not suggest it as a substitute for a lawyer's professional judgment.

What Was Announced on September 17, 2026: Why “Configured” Is Not “New Model”

The official announcement opens by positioning Astra for Law as “our most powerful model, configured into a new AI foundation for law”; the text does not mention training a separate new model, and what changed sits outside the model itself. Even so, it will appear in the model picker as “GPT‑6 Astra Law,” and its API identifier is gpt-6-astra-law.

These additions — a legal search index, a set of tools, and terms governing eligibility and data handling — are explained section by section below. We covered OpenAI's version of ChatGPT for financial institutions a week earlier (event date September 10, 2026); this time it's law firms, with a different index, benchmark, and eligibility conditions, and this article does not repeat what that one already covered.

GPT‑6 Astra is the model the announcement describes as its “latest and most powerful”; we've already covered its capabilities in a separate article (event date September 3, 2026) and won't repeat that here. This article covers only what's added once it's configured into a legal foundation.

What Was Actually Added: a U.S. Legal Index, and Citations You Can Check

The official announcement says Astra for Law can search U.S. case law, statutes, regulations, court rules, and administrative decisions, with an index described as “more than 230 million URLs” and “sources added daily.” Two things need to be kept separate here: the index's scope is explicitly the United States — OpenAI's own text prefixes the entire list with “U.S.” — but which country or region selected law firms may come from is not stated, and the two cannot be inferred from each other.

On case law, OpenAI says it worked with the nonprofit Free Law Project, the organization behind CourtListener, whose case-law collection “covers more than 99.9% of published U.S. precedential case law.” OpenAI also states that this index complements, rather than replaces, the licensed content and specialist products firms already rely on, citing Thomson Reuters as an example; the announcement adds, “Building on OpenAI should mean getting more from the ecosystem, not replacing it.”

Beyond retrieval, the selling point OpenAI emphasizes most is that sources can be checked and opened: OpenAI's own framing is a stronger research foundation, with “reliable authorities the lawyer can examine for herself.” This is the point this article thinks matters most for general readers — not that the model got smarter, but that it now comes with a set of sources you can check yourself.

How to Read OpenAI's Benchmark: 200 Questions, a Private Validation Set, and Matched Reasoning Effort

Every benchmark OpenAI published is a test OpenAI ran itself; we have not independently verified any of it. With both systems at the highest reasoning effort, OpenAI tested Astra for Law's complete setup on 200 U.S. legal research questions drawn from Vals AI's “Legal Research Bench” private validation set. OpenAI says Astra for Law passed the evaluation's overall correctness check on 54.0% of questions, versus 38.7% for GPT‑6 Astra using web search alone, which OpenAI calls a “40% relative improvement.” “Private” means the questions themselves are not public, so outsiders cannot rerun these 200 questions and can only cite the results OpenAI published.

On retrieval, OpenAI separately states that on an audited set of target passages, when comparing the two systems at the same reasoning effort, Astra for Law retrieved “up to 54% more relevant passages from the correct court opinions.” That “up to” is a hedge in OpenAI's own text, meaning not every question sees a 54% gain.

Taken together, what OpenAI is emphasizing with these figures is an improvement in indexing and retrieval, not simply a smarter model; all four percentages hold only under OpenAI's own test conditions, and a different set of questions could produce different results. The two outside quotes in the announcement that directly comment on Astra for Law use phrases like “preview early versions,” “this early stage,” and “early testing” — impressions from an early stage, not formal results.

Who Can Use It and How Data Is Handled: Selected Firms and Two Data Terms

The official announcement draws eligibility narrowly: Astra for Law will be initially offered to selected law firms, through Trusted Access in ChatGPT and Codex, with the API coming soon; it's meant for lawyers at eligible firms and people working under their supervision, for professional legal work. But the announcement does not state the eligibility criteria, firm size, or number of spots, and “coming soon” comes with no timeline. The announcement closes with a single line: to explore early access or build it into your own product, contact OpenAI — beyond that, there's no public application process.

Data terms are handled separately: OpenAI lists two provisions, both scoped to eligible firms — Zero Data Retention on the API, and ChatGPT Enterprise usage excluded from human review by default. OpenAI also mentions it is working with Latham & Watkins, which it describes as a leader in AI governance, to design for information permissions, ethical walls, client instructions, and firm oversight; a separate firm, Wachtell, Lipton, Rosen & Katz, is described as exploring together how AI can support work that requires professional judgment. The announcement frames both as ongoing collaborations, not yet a published standard.

On plugins, OpenAI separately maintains a plugin directory; as read on September 18, 2026, its filter categories include Legal alongside Small Business, Productivity, Engineering & IT, Data & Research, Finance, and others. The announcement says 26 partner-built plugins launched the same day, and a further 9 community plugins written by lawyers and legal engineers add 47 custom skills.

The same announcement separately describes custom applications built with three firms, using the phrase “adapting ChatGPT Enterprise” rather than Astra for Law; Cooley is one of them, with GO Public. Yet Cooley's customer story, published the same day, states plainly that GO Public is built on ChatGPT Work, and notes that in 2025 Cooley advised on 180 deals globally, totaling more than $51.5 billion — a figure about the firm's overall business, which OpenAI does not connect to any AI outcome, and this article does not read the two together either.

What OpenAI Hasn't Made Clear, and the One Thing General Readers Can Use

Across three official pages checked through September 18, 2026, several things go unstated: no price is published; no regional list of eligible firms is given, and whether Taiwanese firms can apply is not addressed; no timeline is given for API availability; and the announcement's body never mentions personal ChatGPT accounts, so it's unclear whether it will ever open to individual users.

The one thing in this announcement more relevant to general readers is that ChatGPT for Word became generally available the same day: it can help proofread, suggest edits, and flag formatting issues, making it easier for lawyers to work inside Word, the tool they already use. But OpenAI states only that it's generally available, without saying which plans, regions, or languages are covered, so this article can only relay the scope OpenAI actually wrote.

Here is an example designed by the editors, not a hands-on test: if a reader receives an AI-generated explanation of a contract clause that cites a statute that doesn't actually exist, the fix this announcement demonstrates isn't a smarter model — it's a retrieval system with labeled sources you can click open and check. The same approach applies to any AI answer you run into day to day: ask first whether you can open the source it cites, then decide whether to trust it.

2026 AI News Roundup: Highlights and Daily Life Applications from January to September

ChatGPT for Financial Services Launches: An AI Tool for Financial Institutions, Not Your Personal Account

• AI tools
• Apps & software
• AI news