OpenAI Shipped a Legal Astra, and Harvey Is One of Its API Customers: Nine Days On, the "Who Owns the Entry Point" Question Has a Layered Answer

OpenAI released Astra for Law on September 17. It is not a new model — it wraps GPT-6 Astra in a legal search index, instructions for legal reasoning and drafting, and settings tuned for professional work. The search index reaches US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs with sources added daily, with the case law supplied by the nonprofit Free Law Project's CourtListener database. The benchmark OpenAI gives is 200 questions from Vals AI's Legal Research Bench, where it passed the overall correctness check 54% of the time against 38.7% for GPT-6 Astra with web search alone; reporting cautions that 200 questions drawn from a third party's private validation set is a small sample and not a general measure of legal accuracy. On availability, it goes first to selected law firms through Trusted Access in ChatGPT and Codex with API access coming soon, appears in the model picker as GPT-6 Astra Law, and targets Am Law 200 firms. Partners include co-development with Sullivan & Cromwell, Ropes & Gray and Cooley, while Harvey and Legora are named as API customers and Wachtell Lipton and Latham & Watkins are among firms with initial Trusted Access.

The Answer to Nine Days Ago Is More Interesting Than Either/Or

Writing about Harvey's $550 million round at a $15.5 billion valuation on September 10, this site posed a question: three years out, does the entry point to legal workflow belong to a vertical vendor like Harvey, or to a legal agent supplied directly by a model lab? The risk flagged then was precisely that OpenAI and Anthropic had started working the legal space directly. Nine days later OpenAI shipped a legal configuration — and in the same announcement, **Harvey is listed as an API customer.** So the answer is not replacement but layering: the model vendor takes the model and the search index; the vertical vendor keeps the workflow, the audit trail and the client relationship. Harvey's round was earmarked mainly for building proprietary models, and that move reads more clearly now — it isn't about surpassing its upstream supplier, it's about not being held by the throat by one. Because it is layered, the advice from that piece matters more now, not less: settle in the contract whether templates, retrieval indexes and audit records can be taken with you. The upstream relationship is now a supplier relationship, and which layer your data lives in — and who it follows — is a question for before you sign.

"Not a New Model" Defines the Capability Boundary

This has to be stated plainly: Astra for Law did not change models. The improvement comes from the layer around it — a professional search index plus instructions for legal reasoning and drafting. So the gap between 54% and 38.7% measures the value of *giving it a professional retrieval source*, not the value of *a model that understands law better*. That distinction is directly useful in procurement: if you already have reliable retrieval over legal corpora, what this product adds on top of that is less than the difference between those two numbers suggests. The 54% needs its terms read too: 200 questions, drawn from a third party's private validation set, run by the vendor. Reporting explicitly cautions that it is not a general measure of legal accuracy. One more thing worth recording is the foundation: the case law comes from the nonprofit Free Law Project's CourtListener. Public data infrastructure carrying a commercial product is a pattern this site also met writing about AlphaGenome Atlas on September 10 — a free public layer as the base, with the commercial layer doing access and distribution.

A Product at Roughly Coin-Flip Accuracy Can Only Ship as "the Buyer Must Check the Work"

Read the 54% literally: on that set of legal research questions, close to half come back wrong. That is not a pejorative — it determines the product's shape. The target customers are Am Law 200 firms, where every citation already gets checked line by line, so the division of labor — model drafts, lawyer verifies — is self-consistent, and the positioning matches the capability. The misreading risk is at the other end: **if anyone expects this to remove the verification step, they are using it wrong.** In legal work, skipping the check does not cost you a rewrite; it costs you a wrong authority. Availability is still narrow: selected law firms first via Trusted Access in ChatGPT and Codex, with API access coming. What an evaluator can usefully do now is assemble their own most typical research questions — once the API opens, running your own set will tell you more than those 200 scores.

via: OpenAI: Introducing Astra for Law, SiliconANGLE, Technology.org