OpenAI has launched Astra for Law, connecting GPT-6 Astra to a purpose-built US legal search index and opening it to law firms and legal-tech companies. The model is labeled GPT-6 Astra Law and is rolling out first as Trusted Access inside ChatGPT and Codex, available to a select group of firms for evaluation; API access follows later, with a zero-data-retention option included. Latham & Watkins, Ropes & Gray, Cooley, and Sullivan & Cromwell are among the first firms named.
The scorecard OpenAI is citing comes from Vals AI's private Legal Research Bench evaluation set. Across 200 US legal research questions, with both systems run at their highest reasoning tier, Astra for Law scored 54.0% overall, compared with 38.7% for GPT-6 Astra relying on web search alone — a roughly 40% relative improvement. On case-law-heavy questions, Astra for Law retrieved 24% more relevant precedents; across a human-audited set of target passages, it surfaced up to 54% more relevant passages at the same reasoning tier.
The Real Change Is the Index
The most significant part of this release is retrieval. OpenAI's own legal search index covers more than 230 million URLs, spanning US case law, statutes, administrative regulations, court rules, and agency decisions, which the company says it updates daily. It has also partnered with the Free Law Project to plug in CourtListener's case database, which holds more than 99.9% of published US precedent. The launch also ships with 26 partner plugins and 47 community plugins, connecting products from Relativity, Clio, iManage, Intapp, DeepJudge, and Thomson Reuters into the workflow.
General-purpose models have long stumbled at the same point in legal research: web search turns up case sources of mixed quality and inconsistent versions, with citation chains that break easily, and the cost of a lawyer's review often exceeds whatever time the model saved. Building the index in-house effectively moves that risk from the lawyer's side to the vendor's. The 54.0% figure also tells the other half of the story — more than half of the questions still came back wrong, and OpenAI itself says lawyers must verify every citation individually, keeping governance processes in place for high-stakes matters.
Selling the Index vs. Selling the Workflow
In the legal AI race, Harvey is already valued at $11 billion, and Thomson Reuters has CoCounsel in hand; for both, the moat has never really been the model itself but rather case context, permissioning, and audit trails. OpenAI is entering at a lower layer this time: it isn't building firm-level project management, but packaging retrieval and reasoning into a base that others can integrate — and Thomson Reuters itself is among the 26 launch-partner plugins.
Thomson Reuters' response was equally direct, stressing that CoCounsel remains a trusted, professional system purpose-built for legal work. Harvey's side praised the retrieval capability; Cooley talked about faster workflows. Put those reactions together and the division of labor becomes fairly clear: capability at the model layer can be sourced externally, but who takes on compliance and liability is still each vendor's own business to sell.
For Now, China's Law Firms Are Watching From Next Door
Astra for Law's index covers only US legal sources — not a single Chinese jurisdiction is included. What domestic firms can borrow in the near term is the shape of the thing, not the product itself: turning court judgments, statute databases, and local regulations into a retrieval layer a model can query directly, and only then talking about accuracy. Some Chinese firms are already working on similar efforts, though the benchmark data made public so far is far thinner, leaving no real basis yet for a side-by-side comparison.
OpenAI has not disclosed pricing, how many firms Trusted Access will cover, or what tier the API will open at. Vals AI's questions are drawn from a private evaluation set that outsiders cannot reproduce. None of the early customers named so far have disclosed their actual usage scale.
Sources: OpenAI official announcement, Vals AI Legal Research Bench, CocoLoop, Artificial Lawyer, Neowin; accuracy figures, index size, and plugin counts checked against OpenAI's published figures.