OpenAI targets law firms with Astra for Law

Ask any lawyer about using general AI for legal research, and you will hear the same nightmare: convincing arguments citing cases that never actually existed.

OpenAI wants to put that fear to rest. On September 17, the company unveiled Astra for Law, a specialized foundation designed specifically for law firms and legal tech vendors, built directly on top of GPT-6 Astra.

A dedicated index covering 99.9% of US case law

Legal work does not start with flowery language. It starts with finding the exact binding precedent, locating the right passage, and understanding how risk shifts across contract terms.

To tackle this, Astra for Law integrates a dedicated legal search index covering more than 230 million URLs across statutes, regulations, court rules, and administrative decisions. Partnering with the nonprofit Free Law Project, it pulls in CourtListener data covering over 99.9% of published US precedential case law.

In testing on Vals AI's Legal Research Bench across 200 US legal questions at maximum reasoning effort, Astra for Law hit an overall correctness score of 54.0%. That is a 40% relative leap over standard GPT-6 Astra using web search alone, which scored 38.7%. On case-law tasks, it surfaced 24% more reference cases and retrieved up to 54% more relevant passages from target court opinions.

Beyond search, the platform adds tailored instructions for legal analysis. It helps attorneys separate binding holdings from dicta, surface adverse authorities that weaken their own arguments, and spot contractual loopholes.

Big Law moves in with custom playbooks

Legal adoption always hinges on confidentiality. Law firms simply cannot risk client data bleeding into training runs.

OpenAI introduced a Trusted Access Program for eligible firms, featuring Zero Data Retention on its API and turning off human review by default in ChatGPT Enterprise. Latham & Watkins assisted in crafting governance controls, including information permissions and ethical walls.

Several elite firms have already put forward-deployed setups to work. Sullivan & Cromwell built an agreement analyzer that incorporates firm negotiation playbooks into deal reviews. Ropes & Gray developed a diligence tool to trace risks across virtual data rooms, while Cooley launched GO Public to streamline IPO filings.

OpenAI also rolled out 26 partner plugins connecting to staple software like iManage, Intapp, DeepJudge, and Thomson Reuters HighQ, alongside 9 community plugins and the general release of ChatGPT for Word. The model will appear in ChatGPT as GPT-6 Astra Law and soon launch on the API, where partners like Harvey and Legora plan to integrate it.

Why it matters

We think OpenAI played this hand smartly. Instead of pretending a general chatbot can practice law, it tackled the two biggest objections from managing partners: bad citations and privacy risks.

For legal professionals, a 54% benchmark correctness rate means human oversight remains mandatory. But for deal diligence and preliminary case research, cutting out hours of manual cross-referencing is an immediate productivity win. It lets junior associates hand off the tedious paper chase to the model while keeping the strategic drafting for themselves.

For the tech industry, OpenAI is positioning itself as the foundational layer of legal tech rather than a standalone app. By integrating with existing platforms like iManage and partnering with Thomson Reuters, it brings traditional software into its orbit. The real pressure lands on niche legal wrappers, who now have to prove their added value when the underlying model already knows the law this well.