OpenAI Astra for Law: AI Legal Research With Real Numbers (2026)

Last updated: September 2026

54.0% versus 38.7% — that single comparison is the whole story of OpenAI's September 17, 2026 announcement. The company launched Astra for Law, a legal-specialized configuration pairing its strongest model GPT-6 Astra with a legal search index covering more than 230 million sources. On an independent validation set of 200 US legal research questions, the complete system passed the overall correctness check on 54.0% of questions versus 38.7% for the same model using ordinary web search — a 40% relative improvement in one step. The question is no longer "can AI do legal research?" but "who adopts the tool first?"

A lawyer in London, a legal-tech founder in Singapore, and an LLM student in Toronto all read those numbers and immediately ask two questions: does it cover my jurisdiction, and can I get access? This article answers both honestly — including the parts you won't like. Major New York firms are already building on it, while the practical path for the rest of the world runs through named partners and specific conditions you should understand today, not next year.

What Astra for Law actually is

Per the official announcement, it is "a new foundation for law firms and legal technology companies to build AI products and workflows around their expertise." The composition has three parts:

• The model: GPT-6 Astra, OpenAI's latest and most capable model.
• The index: a dedicated legal search engine instead of general web search.
• The customization: settings, tools, and context tuned for legal analysis and legal writing — from distinguishing a court's holding from its incidental observations, to engaging with the precedents that weaken your argument rather than only those that support it.

Legal-tech companies Harvey and Legora are building on it through the API from day one, and firms including Sullivan & Cromwell, Ropes & Gray, and Cooley revealed that OpenAI's forward-deployed engineers built custom tools with them inside ChatGPT Enterprise — details below.

The announced numbers, and what they honestly mean

The evaluation ran on the Vals AI Legal Research Bench, a private validation set of 200 US legal research questions measuring whether the system finds the right sources and passages and whether its answers meet the evaluation criteria:

An honest reading requires two clarifications. First, 54% success means 46% of questions still failed — a huge jump, but an assistant for a reviewing lawyer, not a replacement for one. Second, these are US questions judged by US criteria; the immediate beneficiaries are those working with US law, as the limitations section details.

Source: OpenAI's official Astra for Law announcement

How the model searches 230 million sources

The power is not magic but massive information architecture. First, what the index covers: US case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily. The key partnership is with the Free Law Project, the nonprofit behind CourtListener, whose case-law collection — covering more than 99.9% of published US precedential case law — sits inside the research experience.

In practice this means that when a lawyer asks about a precedent, the system does not return "articles discussing the case" but the case itself, with its passages, an understanding of how binding it is, and how it relates to the matter at hand. The announcement is careful to position the index as complementing licensed content and specialist products firms already rely on from providers like Thomson Reuters — integration with the existing ecosystem, not replacement of it.

Source: Vals AI official site

What major firms already built on top of it

Three documented cases from the announcement itself, each a template for a different workflow:

• Sullivan & Cromwell built an agreement analyzer that brings the firm's negotiating playbooks and selected precedents into reviewing a new deal: it surfaces risks that emerge only when provisions are read together, then turns findings into proposed redlines and draft client advice the lawyer can challenge and refine.
• Ropes & Gray built a deal diligence system mirroring how its lawyers work through a data room: it traces every finding back to source and pinpoints questions that could affect an acquisition — like whether key customer contracts require notice or consent.
• Cooley built GO Public to bring its capital-markets expertise into IPO preparation, from drafting the filing to surfacing the risks deserving management's attention — and when the deal changes, it carries that change across the whole filing.

At the deepest layer, a research partnership with Wachtell, Lipton, Rosen & Katz explores supporting sophisticated legal judgment itself — a signal that the platform is targeting the hardest parts of the profession, not just the easiest.

Source: Harvey official site

What this means for law firms worldwide

Here we separate promise from reality without flattery:

• For firms working with US law: this is directly actionable news — firms anywhere serving clients with US investments or disputes can request early access today and benefit immediately from the US precedent index. And this segment is not small: international commercial arbitration, cross-border M&A, and US compliance work for companies worldwide all involve US legal research daily.
• For local-jurisdiction work: not available — the index is thoroughly American, with no announced coverage of English, EU, Commonwealth, or other national law. Anyone selling you "local-law Astra for Law" today is ahead of the truth.
• For the next phase: through companies like Harvey and Legora building on the API, similar capabilities will arrive as finished products — and these companies historically serve firms across multiple regions, making them the natural access channel.
• For law students: the market's message just got louder: the skill of "verified AI-assisted research" — question formulation, source evaluation, spotting the failure — is now part of the definition of a competent lawyer. Students who train it today will lead their firms tomorrow, and Truescho's scholarship and skills resources support exactly that path.

A readiness checklist for your firm

Between enthusiasm for the numbers and responsible adoption stands a plan. This is a practical checklist for any firm reading this news today:

• Audit your talent before your tools: the lawyer who writes precise instructions and evaluates sources will multiply their output; the one who copies the first answer will endanger the firm. One internal training session beats any subscription.
• Inspect your information architecture: upcoming legal AI lives on organized documents. A digitized archive, indexed case files, and structured precedent banks are the fuel that makes any future tool specifically useful to your firm.
• Choose first use cases wisely: start with research, summarization, and assigned agreement review — not final briefs. Every major firm in the announcement began with one well-defined internal workflow and expanded outward.
• Write a data policy now: which client files may ever reach a cloud tool, and which never do? The firms that receive this technology earliest will be the ones whose policy was finished before it arrived.
• Follow Harvey and Legora: they are the most realistic access route outside the US, and when they ship capabilities on top of Astra for Law, you'll be first to know if you're watching their announcements.

Quick comparison: Astra for Law versus today's alternatives

Limitations: what it cannot do — believe nothing more

• US-only index at launch: no Arab, European, UK, or Commonwealth law is announced within the index.
• Access is restricted: initially to selected firms through the Trusted Access program inside ChatGPT and Codex, with the API "coming soon" — you are not yet in a "subscribe here" phase.
• Naming matters: it appears in the model picker as GPT-6 Astra Law and in the API as gpt-6-astra-law — keep this distinction straight so you don't confuse it with the general model.
• Privacy is strong but conditional: Zero Data Retention on the API and exclusion of ChatGPT Enterprise usage from human review by default are serious guarantees, but they attach to the Trusted Access program for eligible firms — not an automatic right for every user.
• 46% failure rate: the number itself says human review is mandatory — the system finds and analyzes precedents; the decision and the responsibility remain yours.
• The ecosystem plugins (26 partner plugins including iManage, Intapp, DeepJudge, and Thomson Reuters HighQ, plus 9 community plugins and 47 custom skills) all orbit the US work environment for now.

FAQ about Astra for Law

What is OpenAI Astra for Law?

A legal-specialized configuration OpenAI launched on September 17, 2026, combining the GPT-6 Astra model with a legal search index covering more than 230 million US sources and tailored instructions for legal analysis and writing, letting law firms and legal-tech companies build products and workflows on top of it.

How accurate is Astra for Law at legal research?

On the Vals AI Legal Research Bench's 200 US questions, it passed the overall correctness check on 54.0% of questions versus 38.7% for the general model with web search, retrieved 24% more case references, and surfaced up to 54% more relevant passages — a major improvement that still requires lawyer review.

Does Astra for Law support non-US law?

No. The announced index covers US case law, statutes, regulations, court rules, and administrative decisions, with the Free Law Project/CourtListener partnership covering US precedents. No other jurisdictions are officially announced in the index so far.

How do I get access to Astra for Law?

Initial access runs through the Trusted Access program for selected firms within ChatGPT and Codex, by contacting OpenAI directly, while API access is "coming soon" per the announcement. For most firms, the faster realistic route may be products from Harvey or Legora built on top.

What's the difference between Astra for Law and Harvey?

Astra for Law is OpenAI's base capability layer (model + index + settings), while Harvey is a legal-tech company building a complete, firm-ready product on top of that layer and others — the first is the engine, the second is the car you drive without building an engine yourself.

Is it suitable for solo lawyers and small firms?

The Trusted Access program targets selected eligible firms at this stage, and the announced custom builds all belong to large firms, so direct access is difficult for solo practices today — and the 54% figure itself makes using it without review professionally risky in any case.

Is my client data safe with it?

For eligible firms in the program: zero data retention on the API, ChatGPT Enterprise usage excluded from human review by default, and permissions and ethical walls designed in partnership with Latham & Watkins — but these guarantees attach to the program, not to ordinary accounts.

The bottom line

Keep the two governing numbers: 230 million sources searched daily, and 54% independent-benchmark success versus 38.7% for its predecessor. Between those two figures lives an entire era of legal work: tools stronger than ever, still requiring lawyer judgment and discipline to become sound professional practice. For everyone outside the US, the play for now is knowledge, not ownership — understand what's happening, train your team on verified research, and watch Harvey and Legora, because when regional coverage arrives, the first contracts will go to those who prepared early. We're watching alongside you, step by step.

Sources

• Official announcement: Introducing Astra for Law — OpenAI — the complete launch post with all figures
• Vals AI — independent legal research benchmark authors — the evaluation standard used
• Free Law Project / CourtListener — the index's information partner
• Our coverage of the GPT-6 Astra launch — the base model
• OpenAI's Agents API launch — the building layer for specialized sectors