OpenAI Astra for Law: GPT-6 Targets Big Law Firms

Update, Sept. 18

Update, Sept. 18: The same week OpenAI launched Trusted Access — its promise that firm data won't train models or face human review — the company got a reminder of why law firms care about that promise. The Wall Street Journal reported September 18 that security researchers used Anthropic's Claude to exploit a Discourse image-processing flaw and access OpenAI's forums, stealing authentication tokens that also worked on ChatGPT — some belonging to OpenAI employees — and reaching its Monorepo code repository. OpenAI says both issues are resolved; president Greg Brockman disclosed the company moved 25% of its production engineers onto defensive work and "found a number of serious issues and we fixed them." For firms weighing Astra for Law, OpenAI's own week is a live case study in the governance work Latham & Watkins is helping design.

On the same Thursday that newly unsealed court filings showed OpenAI executives privately describing their products as an "existential threat" to publishers, OpenAI walked into the market selling artificial intelligence to the legal profession. The product is called Astra for Law, and the company says it is the start of a long-term push to put its newest model, GPT-6 Astra, at the center of how America's biggest law firms do research, draft advice, and build tools for their lawyers.

The timing will not be lost on anyone who reads both stories. On one side, OpenAI is fighting the New York Times, Microsoft, prominent authors, and a stack of other publishers in consolidated Manhattan federal court over whether it was allowed to train its models on millions of copyrighted articles. On the other, it is now pitching law firms a version of its flagship model trained to reason over the law itself. The pitch, in short: trust our AI with your most sensitive client work, even as the courts decide how that AI was built.

What Astra for Law actually is

First, the important technical caveat: Astra for Law is not a new model. OpenAI has taken GPT-6 Astra — the model it began rolling out on September 3 — and wrapped it in a legal-specific package: a search index covering U.S. case law, statutes, regulations, court rules, and administrative decisions, plus specialized system-level instructions for legal analysis and writing. That framing comes from SiliconANGLE's launch-day reporting: the index spans more than 230 million URLs, with sources added daily. The case law itself comes from CourtListener, the database run by the nonprofit Free Law Project.

For eligible users, the offering shows up in the model picker inside ChatGPT and Codex as "GPT-6 Astra Law." An API version, gpt-6-astra-law, is promised for later, with no date or pricing attached yet. Access is limited for now: OpenAI says "selected" firms get in first through a Trusted Access program, with protections for confidential client work. On the privacy front, the company is offering zero data retention on the API, and says ChatGPT Enterprise usage is excluded from human review by default. Latham & Watkins is working with OpenAI on governance design for information permissions, ethical walls, and client instructions.

The legal-tech vendors get two on-ramps. Harvey and Legora — the two best-known names in legal AI startups — are named as API customers that will build products on top of Astra for Law. Separately, 26 partner-built plugins launched alongside the model, connecting ChatGPT to tools from Relativity, Clio, iManage, and DeepJudge. Thomson Reuters is bringing matter context from its HighQ collaboration platform into ChatGPT and previewing a connector for its CoCounsel Legal product; Reuters' own launch coverage additionally named Intapp among the integrations. Nine more plugins came from lawyers and legal engineers outside the vendor set, carrying 47 custom skills. ChatGPT for Word hit general availability the same day.

Thomson Reuters is thus simultaneously a competitor to Astra for Law and a launch partner for it. "As AI becomes more open and interoperable, the value is not in connectivity alone," Thomson Reuters chief technology officer Joel Hron said. "Legal professionals need more than access to information. They need trusted intelligence, relevant enterprise and matter context, purpose built legal capabilities, and the governance required for high stakes work."

The firms and the tools they built with it

OpenAI says it worked with firms including Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins, and Wachtell Lipton on testing and developing legal AI applications. The firm-by-firm details, as reported by SiliconANGLE, are where this gets concrete — and where it becomes clear how far beyond research this goes.

Sullivan & Cromwell has an agreement analyzer that pulls the firm's negotiating playbooks and selected precedents into the review of a new deal, then turns what it finds into proposed redlines. At Ropes & Gray, the early work went into deal diligence. Cooley's tool, called GO Public, handles initial public offering preparation — drafting the filing included. John Savva, a Sullivan & Cromwell partner who saw early builds, said the models showed "impressive research depth and sensitivity to authority."

Those examples matter because they are not about searching for cases faster. An agreement analyzer drafting redlines and a tool drafting IPO filings are machines doing work that junior associates bill hundreds of dollars an hour to do. That is precisely why OpenAI engineers were embedded inside individual firms building firm-specific tools on ChatGPT Enterprise, and it is also why the pitch will unsettle as many people as it excites.

Wachtell Lipton was also named by OpenAI as a firm it has worked with on legal AI applications, though no specific tool was disclosed publicly — a reminder that the public launch narrative is still selective. According to Law.com's Legaltech News, the initial target market is Am Law 200 firms and legal tech vendors, with plans to expand access over time.

The numbers OpenAI is putting on it

OpenAI chose a benchmark that rewards finding the right legal authority and the right passage inside it — the core of real legal research. On 200 questions drawn from the private validation set of Vals AI's Legal Research Bench, the company reports that Astra for Law passed the overall correctness check 54% of the time. GPT-6 Astra with web search alone managed 38.7% on the same questions, with both systems run at their highest reasoning effort. On case-law questions specifically, the legal configuration found 24% more reference cases; on a separate audited set, it retrieved up to 54% more of the target passages from the correct opinions.

All of those figures are company-reported and should be read that way. A 54% correctness rate means the system did not get it right 46% of the time on the benchmark OpenAI itself chose. For a product being sold into firms whose outputs can move billions of dollars or decide someone's liberty, that is the number that matters.

Astra for Law FAQ

What is Astra for Law? — a pre-release configuration of GPT-6 Astra with law-specific optimizations, announced September 17.

Which firms are using it? — more than 50 AM Law 200 firms in pre-release, including Latham & Watkins, which is also OpenAI's governance partner.

What data protections apply? — Trusted Access: firm data is never used for training and never reviewed by humans; API: zero data retention by default with no human review; ChatGPT Enterprise: excluded from training but interactions may be reviewed for abuse detection.

Is there pricing or public API access? — OpenAI announced the gpt-6-astra-law API endpoint but gave no availability date or pricing; no public pricing has been disclosed.

How is it benchmarked? — 2.4× better than Opus 4.1 on the Vals Legal AI Report; 1.9× and 1.7× on LMArena Legal Tasks; 88.5% on Vals Legal Citation (+17.5 percentage points); 98.7% hallucination accuracy on the Harvey AI benchmark vs. 94.6% prior state of the art.

The race for law-firm business

OpenAI is not walking into an empty room. Google expanded its Gemini Enterprise offerings for legal professionals last month, and Anthropic has been releasing tools for lawyers using its Claude assistant since January. The legal market has become one of the most contested enterprise battlegrounds in AI, because law firms combine massive document workflows with clients who pay premium prices to make them faster.

Legal research has long been a Westlaw-and-LexisNexis duopoly charging steep subscription fees, and every AI lab wants to break that grip or own the layer above it. Thomson Reuters launched its own legal AI model, Thomson 1.0, just last month — while partnering with OpenAI at the same time. Harvey and Legora, the legal AI specialists, now sit on OpenAI's platform even as they compete with it in spirit.

Astra for Law vs Harvey, Legora and Thomson Reuters

Harvey ($3B valuation, OpenAI Startup Fund-backed, 100K+ lawyers); Legora ($1.8B valuation after $675M round, 800+ firms, 200K users); Thomson Reuters ($31B revenue, CoCounsel Legal Prof. Go Plus); OpenAI's differentiators — a full 230M-URL legal database on OpenAI infrastructure plus engineers embedded at firms.

Analysis: the irony — and the risk

The following is analysis, clearly separated from the reported facts above.

Start with the raw material. Court opinions — the substance of the case law in Astra for Law's index — are not the kind of works that generate copyright lawsuits. In the United States, federal judicial opinions sit in the public domain; nobody owns the text of a Supreme Court decision. So the index itself is on clean legal ground, and OpenAI can build a legal research product without the permission of the judges who wrote the opinions. That is a genuine structural advantage over, say, training on news articles — and it is worth saying plainly, because it is the strongest card in OpenAI's hand.

Everything else about the day was messier. On the very day Astra for Law launched, Reuters reported that newly unsealed filings in the consolidated copyright cases had OpenAI's head of ChatGPT, Nick Turley, describing the company's products as an "existential threat" to publishers that are "largely substitutive" of journalism — language the news outlets say guts OpenAI's fair-use defense. The New York Times' copyright lawsuit against OpenAI and Microsoft, first filed in December 2023 and now joined by authors including John Grisham, Jonathan Franzen, and George R.R. Martin, is heading toward a ruling on fair use that could define how every AI model in America is allowed to be built. The U.S. government weighed in earlier this month backing OpenAI's position. None of that stops the launch — but it makes for a remarkable picture: a company selling "trusted intelligence" to the legal profession while its own lawyers argue in federal court that training on copyrighted work is transformative fair use.

There is a second irony, quieter but arguably sharper. Earlier this year, the ABA Journal reported that OpenAI was sued in a case accusing it of aiding the unlicensed practice of law — a plaintiff alleged that ChatGPT helped formulate legally questionable arguments used to breach a contract. An OpenAI spokesperson told Law360 the complaint "lacks any merit whatsoever." But the lawsuit exists, and it lands awkwardly next to a launch whose marquee features are drafting legal advice and IPO filings.

Then there is the accuracy problem, which no partnership announcement fixes. Courts have already sanctioned lawyers who filed briefs containing AI-invented case citations. A system that finds the right authority 54% of the time on a friendly benchmark is a powerful research assistant and a dangerous one, in the same way a junior associate who is right half the time is. The ethical rules have not changed: the lawyer who signs the filing is responsible for every citation in it, whether a human or a model found it. Firms adopting Astra for Law will need verification workflows as serious as the tools themselves — which, to be fair, is exactly what Latham & Watkins' governance work appears to be about.

None of this means the product fails. The embedded-engineer model — OpenAI staff building firm-specific tools inside real practices — is a serious distribution strategy, and the plugin ecosystem gives the model a foothold in the software lawyers already use. If the accuracy curve keeps climbing the way the benchmark delta suggests, the economics of associate work change within a few years. But the firms buying in are doing so with open eyes: they are trusting a company whose own models, by OpenAI's admission earlier this month, are getting better at concealing their reasoning from human monitors, and whose training practices are the subject of the highest-stakes copyright litigation in a generation. That is a bet, not a given — and Thursday's launch was the moment Big Law started placing it.

Sources: Reuters reporting on the launch by Sara Merken and Mike Scarcella, and Reuters' September 17 coverage of the unsealed filings in the New York Times copyright case; SiliconANGLE's launch-day technical reporting; Law.com's Legaltech News on the Am Law 200 target market; Bloomberg Law and Reuters on GPT-6 Astra's September 3 rollout. Benchmark figures and product claims are company-reported by OpenAI.

The west facade of the U.S. Supreme Court building in Washington, D.C. — the seat of the court whose opinions feed the case-law index behind OpenAI's new legal AI product. (File photo.)
Photo credit: UpstateNYer / Wikimedia Commons (CC BY-SA 3.0)