OpenAI unveiled Astra for Law on September 18, a version of GPT-6 Astra built for legal work and backed by an index of more than 230 million URLs of US law. On a legal research benchmark, Astra for Law answered 54% of questions correctly, against 38.7% for GPT-6 Astra running a plain web search.
Key Takeaways
- Astra for Law pairs GPT-6 Astra with legal analysis and writing instructions and a new legal search index refreshed every day.
- Latham & Watkins, Ropes & Gray, Cooley and Sullivan & Cromwell are among the first firms on board.
- Access starts through Trusted Access in ChatGPT and Codex, with the API coming later.
Have an AI Sum Up This Article
ChatGPTWhat Astra for Law Adds on Top of GPT-6 Astra
Astra for Law runs on a model that already exists. OpenAI’s September 18 announcement describes a bundle: GPT-6 Astra, dedicated instructions for legal analysis and drafting, settings tuned for thorough work, and above all a search index built for the law.
That index, called the Legal Search Index, covers US case law, statutes, regulations, court rules and administrative decisions. It spans more than 230 million URLs, and new sources land in it every day.
The collection draws on data from the Free Law Project, which covers more than 99.9% of published US precedents. It’s the piece the general-purpose model launched in early September under an AGI banner was missing: a closed, checkable corpus to search instead of the open web.
Astra for Law also ships with an ecosystem. OpenAI is launching 26 partner-built plugins, from Thomson Reuters, Intapp, Harvey, Legora, DeepJudge and iManage among others, plus 47 community plugins. Several of them hook into software firms already run, such as Relativity and Clio.
That plugin list matters as much as the model. A legal assistant only earns a place in a firm if it can reach the documents, matters and review platforms lawyers already live in, and OpenAI chose to show up with a long list of those connections on launch day, ready to switch on.
The approach extends the Astra agents OpenAI has been showing off in recent weeks, like the one that runs a shop and flies a drone on its own. OpenAI is now slicing the same model by profession, each version with its own tools and context.
Law is a logical place to start the series. The right answer sits in identifiable texts, and a made-up citation costs a lot to whoever repeats it in front of a judge.
A closed index is a direct response to that risk. A model searching the open web mixes rulings with blog commentary and press coverage, while a base limited to official texts lets a lawyer trace every citation back to the decision and read it. The promised gain comes first from the quality of what the model reads.
Astra for Law by the Numbers, as Measured by OpenAI
OpenAI ran its own test on Vals AI’s Legal Research Bench, a set of 200 legal research questions. Astra for Law passed 54% of them, compared with 38.7% for GPT-6 Astra using ordinary web search.
The company flags three more gaps against the same model without the index: answers 40% more accurate, 24% more reference cases found and 54% more relevant passages pulled from the correct court opinions. These figures come from OpenAI alone, and no independent test has been published yet.
Read the other way, a 54% score also means nearly one question in two on the benchmark still went wrong. For a firm, the tool works as a research accelerator whose every output still has to be checked by a lawyer before it goes into advice.
On privacy, eligible firms get zero data retention. It mirrors the way OpenAI watches for abuse without keeping customer data, a key selling point for lawyers bound by client confidentiality.
Access stays gated for now. Astra for Law goes first to selected firms through Trusted Access in ChatGPT and Codex, and OpenAI says it will reach the API soon. Latham & Watkins, Ropes & Gray, Cooley and Sullivan & Cromwell are named as early adopters.
Those two doors point at two kinds of users. ChatGPT serves the lawyer who asks a question and reviews the answer, while Codex speaks to the technical teams inside firms and vendors, who can wire their own pipelines around the index. OpenAI covers the individual desk and the in-house tooling of large firms in one move.
The release pace matters too. Two weeks after GPT-6 Astra, OpenAI already has a sector version out, in line with the six months of roadmap Astra’s productivity bought the company.
More articles on Horizon
- Claude Code Projects Splits Coding Work Across Agents
- Microsoft Exec Called AI Training a Theft of Labor
- Gemini 3.8 Live Test: What Google’s Voice Model Delivers
Harvey and Legora, Customers and Rivals of Astra for Law
For firms, Astra for Law reopens the tooling question. A partner can now get sourced research inside the ChatGPT window they already use, without a specialist legal product in between. The value moves to checking citations, which stays the lawyer’s job.
For specialist vendors, the picture is murkier. OpenAI presents Harvey and Legora as API customers who can build on Astra for Law, even though both startups sell AI-assisted legal research to exactly the same firms.
The money at stake is real. Legora hit a 5.6 billion dollar valuation this spring, in a head-to-head race with Harvey to equip the big firms. If Astra for Law becomes the default research layer, both will have to win on workflows, matter management and integration with in-house tools.
There is a quieter risk for them too. Every query their customers run through OpenAI’s index is a query that no longer depends on their own research layer, and Harvey, which counted OpenAI among its early backers, now watches the model maker sell a finished product into its own client base.
Harvey’s head of research praised the depth of the research and its sensitivity to the hierarchy of legal authority in early testing, while framing Harvey’s own platform as complementary. Thomson Reuters’ chief technology officer pushed back on the idea that connectivity is enough, arguing that legal professionals need trusted intelligence, matter context and purpose-built legal capabilities.
The move lands in a fight already under way. Anthropic stepped into legal services this spring, and OpenAI keeps going with its push to close the enterprise gap with Anthropic. Legal work is turning into a market where AI labs sell both the engine and the application.
What comes next hinges on two things: the API opening, which will show whether Harvey and Legora build on Astra for Law or keep their distance, and independent tests that can confirm the 54% OpenAI is claiming.
Follow the story on Horizon.
