OpenAI launched Astra for Law on September 17, a configuration of its GPT-6 Astra model aimed at law firms and legal technology vendors. The company was explicit in its Astra for Law announcement that no new model was trained: what ships is a legal search index spanning more than 230 million URLs, domain-specific instructions, response settings and tooling wrapped around the existing frontier model.
Key takeaways
- On 200 questions from the private validation set of Vals AI's Legal Research Bench, Astra for Law cleared the overall correctness check on 54%, against 38.7% for GPT-6 Astra with generic web search β 15.3 points, or a 40% relative gain.
- The Legal Search Index covers US case law, statutes, regulations, court rules and administrative decisions, drawing on CourtListener from the nonprofit Free Law Project.
- Access runs through a Trusted Access program aimed at Am Law 200 firms, with zero data retention on the API and ChatGPT Enterprise usage excluded from human review.
What the benchmark numbers actually show
The headline comparison is a configuration against itself. At the highest reasoning effort for both systems, the legal-indexed setup passed correctness on 54% of questions versus 38.7% for the same model using ordinary web search. On case-law questions it surfaced 24% more reference cases, and retrieved up to 54% more relevant passages from the correct opinions at matched reasoning effort. Answers also ran roughly twice as long on average.
That framing is useful and limited in the same breath. It isolates the value of grounded retrieval over open web search, which is the honest thing to measure. It says nothing about how Astra for Law compares with Harvey, Legora or CoCounsel on the same benchmark, and a 54% correctness rate still means roughly one in two research answers failed the evaluation criteria.
Why the index is the product
Jason Boehmig, the Ironclad cofounder OpenAI hired in June to lead its legal products, told reporters at a briefing covered by LawSites that this is a model configuration rather than a new model β domain instructions for legal writing, controls such as response length, and legal-specific tooling that the company expects to expand. Sherwin Wu, the engineer leading the product, demonstrated drafting a motion to dismiss for a hospital facing a disability-discrimination claim, a stockholder challenge to a rushed asset sale, and a sales-commitment dispute.
The strategic point is that firms and vendors no longer have to assemble and maintain the underlying legal corpus themselves. That is a direct move on the layer where legal AI startups have been differentiating, and it lands as those same startups become customers: Harvey and Legora are named as API customers of the thing that now ships their retrieval layer as a platform primitive.
Plugins, partners and the firms building on top
OpenAI paired the launch with 26 partner plugins from vendors including Thomson Reuters, Harvey, Legora and iManage, nine community plugins from legal engineers, and 47 custom skills the company says were built by legal power users rather than by OpenAI. ChatGPT for Word also reached general availability, putting proofreading and edit suggestions inside the document where most legal drafting actually happens.
Several firms have already built on ChatGPT Enterprise with OpenAI engineers embedded on site. Sullivan & Cromwell built an agreement analyzer running on its own negotiating playbooks and precedents; Ropes & Gray built M&A diligence tooling shaped around how its lawyers work a data room; Cooley built a capital-markets system for IPO preparation. Latham & Watkins is working with OpenAI on information permissions, ethical walls and firm oversight.
Read alongside the benchmark, the plugin roster tells you where OpenAI thinks the remaining value sits. Retrieval is being commoditised into the platform; matter context, document management and billing hooks are being left to partners. Vendors that built their differentiation on having the better legal corpus now have to relocate it somewhere the model maker is not standing.
The part firms will actually negotiate
Asked whether the zero data retention promise carried exceptions, Boehmig said the one-sentence version summarises an access agreement running about 30 pages, and that it is more complicated than the sentence suggests while still meeting the need. For general counsel evaluating this, that answer is the story β professional-responsibility obligations attach to the contract, not the marketing page. A firm cannot discharge a duty of confidentiality by citing a product page, and thirty pages is where the carve-outs live.
The competitive backdrop is Anthropic's earlier Claude for Legal push, and the pattern is now familiar: take a frontier model, bolt on a curated corpus and domain instructions, and sell it as a vertical. OpenAI ran the same play in finance when it shipped a Wall Street configuration of ChatGPT. Expect the next few to arrive the same way, because the expensive part is no longer the model.
FAQ
Is Astra for Law a new OpenAI model?
No. OpenAI says it is a configuration of GPT-6 Astra, combining the existing model with a dedicated legal search index, domain-specific instructions and legal tooling. Boehmig made the distinction explicitly during the press briefing.
Who can use Astra for Law right now?
It is coming to API customers including Harvey and Legora, and is offered to selected law firms through a Trusted Access program targeting Am Law 200 firms, where it appears in the ChatGPT and Codex model picker. It is not generally available to individual practitioners.
Where does the legal search index come from?
The corpus spans more than 230 million URLs of US case law, statutes, regulations, court rules and administrative decisions, with new sources added on an ongoing basis. OpenAI says it drew on CourtListener, the research library maintained by the nonprofit Free Law Project.
