OpenAI introduced Astra for Law on September 17, 2026, according to a same day post on its own site: a new AI foundation that combines GPT-6 Astra, the company's flagship model, with a dedicated legal search index, custom instructions for legal analysis, and a Trusted Access Program built around confidentiality controls. Harvey and Legora are named as API partners that will build on top of it, and OpenAI says a group of major law firms, including Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins, and Wachtell Lipton Rosen & Katz, have already worked with its forward deployed engineers on custom tools.

The pitch is simple: pair a frontier model with a legal specific research layer and enough governance controls that firms will trust it with client confidential work. OpenAI backs that pitch with numbers from its own testing, not an independent audit, showing meaningful gains over the base GPT-6 Astra model on a legal research benchmark.

Worth flagging before any of the marketing language sinks in: several of the firms praising the partnership in OpenAI's own post are also currently doing paid legal work for OpenAI itself, and the model underneath this new legal tool is the same one that drew safety scrutiny three weeks ago at its own launch.

Quick Take

  • OpenAI introduced Astra for Law on September 17, 2026, combining GPT-6 Astra with a legal search index covering more than 230 million URLs and instructions tuned for legal research and drafting, offered first through a Trusted Access Program in ChatGPT and Codex.
  • On OpenAI's own 200 question private benchmark, Astra for Law passed 54.0% of questions versus 38.7% for base GPT-6 Astra using web search alone, a result OpenAI reported itself that no outside group has reproduced.
  • Three of the six firms OpenAI names as build partners, Sullivan & Cromwell, Cooley, and Wachtell Lipton Rosen & Katz, have also done significant paid legal work for OpenAI itself, including preparing the company for its own IPO.

Background: How OpenAI Got Here

This launch did not come out of nowhere. OpenAI hired Ironclad co-founder Jason Boehmig in June 2026 to run a dedicated legal industry vertical, a signal the company was building toward exactly this kind of product months before today's announcement, according to Legal IT Insider's reporting at the time.

Then came the model itself. GPT-6 Astra launched on September 3, 2026, and OpenAI called it its most capable release yet. Coverage at the time focused on Astra's use of a reasoning technique called opaque recurrence, which BriefFlash has covered in detail, and on OpenAI's own materials reportedly flagging Astra as the first model to cross the company's "Critical" cybersecurity capability threshold, a detail we reported here two weeks ago.

Legal AI was already a crowded, fast moving market before today. Harvey, which OpenAI seeded early on according to public reporting, and Legora, a Stockholm founded rival that has raised over $200 million and now serves roughly 750 firms, have spent the past year racing each other on features and law school partnerships. Anthropic has its own legal plugin for Claude, and according to industry coverage, its introduction previously moved the stock prices of publicly traded legal software companies. Astra for Law puts OpenAI's own model directly into that fight, as both a supplier to Harvey and Legora and, through its custom firm builds, a potential competitor to them.

The Current Picture

What Astra for Law Actually Combines

Astra for Law is not a new model. It is GPT-6 Astra paired with a legal search index and a set of custom instructions for legal analysis and writing, according to OpenAI's announcement. The index searches U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily, and it incorporates Free Law Project's CourtListener collection, which OpenAI says covers more than 99.9% of published U.S. precedential case law.

OpenAI tested the combined system on 200 U.S. legal research questions drawn from the private validation set of Vals AI's Legal Research Bench. At the highest reasoning setting for both systems, 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, a 40% relative improvement by OpenAI's own math. On case law focused questions specifically, Astra for Law surfaced 24% more reference cases and retrieved up to 54% more relevant passages from the correct court opinions than the base model.

Astra for Law will appear in ChatGPT's model picker as "GPT-6 Astra Law" and in the API as gpt-6-astra-law, with API access for Harvey and Legora coming and general API availability described only as coming soon.

Legal-Grade Trust and Controls

Handling client confidential material means governance, and OpenAI is leaning on that here. Eligible firms in the Trusted Access Program get Zero Data Retention on the API, and ChatGPT Enterprise usage is excluded from human review by default. OpenAI says it is working with Latham & Watkins to design information permissions, ethical walls, client instructions, and firm oversight tools, a collaboration Latham's Michael Rubin described as building on the firm's existing investments in AI governance.

Custom Builds and the Plugin Ecosystem

Beyond the model itself, OpenAI's forward deployed engineers have been building bespoke tools directly with individual firms. Sullivan & Cromwell built an agreement analyzer that folds the firm's own negotiating playbooks into deal review and drafts redlines. Ropes & Gray built a diligence system for working through data rooms on acquisitions. Cooley built a tool called GO Public for IPO preparation that tracks how a filing's changes ripple across a document.

Separately, OpenAI is launching 26 partner built plugins connecting ChatGPT to tools firms already use, including iManage, Intapp, DeepJudge, Relativity, and Clio, plus a Thomson Reuters integration that brings HighQ matter context into ChatGPT and previews a coming CoCounsel Legal connector. Thomson Reuters Chief Technology Officer Joel Hron framed the work as pairing open interoperability with the governance high stakes legal work demands. Nine additional community plugins and 47 custom skills come from legal engineers at LegalQuants, LECG, and Skills.law. OpenAI also made ChatGPT for Word generally available today.

Reality Check

Start with the benchmark. A 54.0% pass rate against 38.7% is a real gap, but it is OpenAI's own test, on a private validation set, scored by OpenAI, and published by OpenAI. Nobody outside the company has reproduced it yet, and Vals AI's role as the benchmark's creator does not amount to an independent audit of this specific result.

Then there is the competitor comparison tucked into OpenAI's own post: a side by side legal memo example where Astra for Law's output is measured against Anthropic's Claude Fable 5.1, with OpenAI's material claiming the Anthropic model cited a precedent that had actually been reversed on appeal in one scenario and failed to find a matching case in the other. That is a single, company selected pair of outputs from the vendor selling the winning product, not a scored, repeatable evaluation, and it fits a pattern this outlet has already tracked this month of OpenAI pairing autonomy and performance claims with customer stories that read more like marketing copy than audited results.

There is also the model underneath all of this. GPT-6 Astra is the same system OpenAI's own materials reportedly flagged as crossing a "Critical" cybersecurity capability threshold, and its opaque recurrence reasoning technique has already drawn scrutiny over how observable its decision making actually is. Handing that model discretion over confidential legal work, even with Zero Data Retention promises, is a different risk calculation than using it to summarize a public contract template.

Finally, look at who is in the room. Cooley and Wachtell Lipton Rosen & Katz, both quoted in OpenAI's announcement praising the collaboration, are the same two firms Bloomberg Law reported OpenAI retained in March 2026 to prepare for its own initial public offering. Sullivan & Cromwell, also featured, has separately advertised advising OpenAI on its multibillion dollar partnership with Microsoft. None of that makes their product feedback dishonest. It does mean three of OpenAI's six flagship legal partners already have an existing, paid relationship with the company whose product they are endorsing, which is worth knowing before taking the endorsements at face value.

Who This Affects

Large law firms with the budget for a bespoke build, the kind Sullivan & Cromwell, Ropes & Gray, and Cooley just got, stand to gain the most immediately, while smaller firms wait for the plugin ecosystem and eventual API access to trickle down.

Legal tech vendors like Harvey and Legora get a more capable underlying model to build on, but also a reminder that OpenAI's own engineers are now building custom tools directly for the same large firm clients those vendors compete for.

In house legal and compliance teams evaluating any of this should treat the Trusted Access Program's confidentiality claims as a starting point for due diligence, not a substitute for it, particularly given the underlying model's own safety profile.

What to Watch

Watch for independent researchers or rival vendors to test Astra for Law's research claims against the base model themselves, since OpenAI has not published enough detail for anyone outside the company to reproduce the Vals AI numbers directly. Watch for whether Harvey or Legora respond publicly to being positioned as both partners and, through OpenAI's direct firm builds, competitors. And watch whether wider API access actually arrives on the timeline OpenAI implies, given the company's pattern this month of announcing features as broadly available before infrastructure catches up, including its own recent pause on new ChatGPT Pro sign-ups after Astra demand outstripped capacity.

Key Takeaways

  • OpenAI introduced Astra for Law on September 17, 2026: GPT-6 Astra combined with a legal search index and Trusted Access controls, initially available through ChatGPT and Codex for selected law firms.
  • On OpenAI's own private 200-question benchmark, Astra for Law passed 54.0% of questions versus 38.7% for base GPT-6 Astra using web search alone, a result that has not been independently verified.
  • Sullivan & Cromwell, Cooley, and Wachtell Lipton Rosen & Katz are named as build partners and quoted praising the launch, and all three also have existing paid legal work for OpenAI, including its IPO preparation.
  • The underlying GPT-6 Astra model is the same one OpenAI's own materials reportedly flagged as crossing a 'Critical' cybersecurity threshold at its September 3 launch.

FAQ

Is Astra a free app?

No. In this announcement, 'Astra' refers to GPT-6 Astra, OpenAI's flagship model, and to Astra for Law, a legal specific configuration of it, not a standalone app. Using it requires a ChatGPT Pro, Plus, Business, or Enterprise plan, or API access, and Astra for Law itself is currently limited to a Trusted Access Program for selected law firms rather than being generally available. This is a separate product from Google's Project Astra, an unrelated live AI assistant project with a similar name.

What is Astra AI?

In the context of this story, Astra refers to GPT-6 Astra, the model OpenAI launched on September 3, 2026 as its most capable release to date, and to Astra for Law, the September 17, 2026 configuration of that model built for legal research and drafting with partners including Harvey and Legora. The name overlaps with, but is unrelated to, Google's separate Project Astra initiative.

What is Astra?

Astra is OpenAI's name for its current flagship model family, GPT-6 Astra, launched September 3, 2026. Astra for Law, announced September 17, 2026, pairs that model with a legal research index and law-specific controls for firms and legal technology companies to build on.

How does Astra work?

For Astra for Law specifically, GPT-6 Astra draws on a legal search index covering more than 230 million URLs of case law, statutes, and regulations, plus Free Law Project's CourtListener collection, then applies custom instructions for legal analysis and writing to turn that research into arguments, drafts, or risk assessments a lawyer can review. Firms can extend it further with 26 new ecosystem plugins connecting it to tools like Relativity, Clio, and iManage.