OpenAI did not ship a new model for lawyers. On September 17, 2026 it introduced Astra for Law, which is GPT-6 Astra, the model it already had, paired with legal instructions and a search index covering more than 230 million web pages of United States case law, statutes, regulations, court rules and administrative decisions. The same launch brings 26 partner-built plugins that connect ChatGPT to the software firms already pay for, so what changes is less the quality of an answer than the place a lawyer starts asking the question.
The model is unchanged, the index is new
OpenAI describes Astra for Law as GPT-6 Astra combined with settings, tools and context tailored to professional legal work, not as a separate base model. Because the underlying model is the same one OpenAI already sells, the difference a lawyer notices comes from two additions: instructions that shape how the system researches and cites, and a corpus it can search that is not the open web. It appears in ChatGPT's model picker as "GPT-6 Astra Law" and, when the API opens, as gpt-6-astra-law.
What is in that corpus matters more than its size. OpenAI says its work with Free Law Project, the nonprofit behind CourtListener, brings a case-law collection covering more than 99.9 percent of published United States precedential case law into the research experience. Precedential decisions are the published ones that later courts are expected to follow, which is why a near-complete set of them is useful in a way that a large pile of legal commentary is not. That collection supplies the case law; the wider index that also reaches statutes, regulations, court rules and administrative decisions is OpenAI's own, with sources added daily. OpenAI says the index is meant to complement the licensed databases and specialist products firms already buy, naming Thomson Reuters among those providers.
The benchmark OpenAI ran itself
On 200 questions from a private validation set of Vals AI's Legal Research Bench, at the highest reasoning effort for both systems, Astra for Law passed the evaluation's overall correctness check on 54.0 percent of questions against 38.7 percent for GPT-6 Astra using web search alone. OpenAI describes that as a 40 percent relative improvement, and it ran the evaluation itself, on questions that are not public. OpenAI reports that the legal configuration also found 24 percent more reference cases on case-law-focused questions and retrieved up to 54 percent more relevant passages from the correct court opinions in an audited set.
OpenAI's own test found the legal setup correct a little over half the time, against under two in five for the same model with ordinary web search. It is also a measurement by the company selling the product, which shows the index adds something without establishing how often a firm would get a defensible answer on a live matter.
The specialist tools arrive inside ChatGPT
The launch includes 26 partner-built plugins connecting ChatGPT to tools including Relativity, Clio, iManage, Intapp, DeepJudge and Thomson Reuters, plus 9 community plugins built by lawyers and legal engineers, carrying 47 custom skills between them. Harvey and Legora, which sell their own legal AI systems, are named as API customers that will be able to build on Astra for Law inside their own products.
What a plugin does is narrower than the name suggests: it lets ChatGPT call another company's system while answering, so matter context or a document set can be pulled in without a lawyer copying text between applications. OpenAI presents the plugins as connections to the specialist tools firms already use, which is why the same launch is additive for those vendors and, at the same time, puts their tools one step further from the moment a researcher starts a question.
Thomson Reuters is bringing HighQ matter context into ChatGPT and previewing a CoCounsel Legal connector that is forthcoming, not live at launch. Its chief technology officer, Joel Hron, described the value as trusted intelligence and the governance required for high stakes work rather than connectivity alone. Firms that tested early versions include Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins, Wachtell Lipton and Skadden, several of which built tools of their own: an agreement analyzer, a deal diligence system, a tool for public offerings. John Savva, a partner at Sullivan & Cromwell, said the models demonstrated impressive research depth and sensitivity to authority across litigation and transactional matters.
Who gets in, and what happens to the files
Access starts narrow. OpenAI says Astra for Law is initially offered to selected law firms through Trusted Access in ChatGPT and Codex, with the API "coming soon." Some early coverage described API access as part of the launch; the announcement lists only Trusted Access as available now. For eligible firms, OpenAI says the offering includes Zero Data Retention on the API, meaning requests are not kept, and that ChatGPT Enterprise usage is excluded from human review by default. Both matter in a profession where a client's confidences are the product, since material a system stores is material that can surface in discovery or in a leak.
What the launch changes
OpenAI supplies the model, the instructions and the index; the specialist tools now arrive as plugins that ChatGPT operates. A firm with access that already pays for Clio for its matters or Thomson Reuters for its research can start the same work inside ChatGPT, because the entry point and the core legal corpus both sit there. That makes where research begins a placement decision rather than a habit set by which suite the firm bought, and it is the part of the launch that the benchmark numbers do not measure.
The next step is the API, which OpenAI says is coming soon. Secondary coverage of the announcement notes that no launch date or pricing has been published. Until it opens, only the selected firms in the Trusted Access program can use the configuration at all.