“As AI becomes more open and interoperable, the value lies not just in connectivity,” said Joel Hron, chief technology officer at Thomson Reuters. “Legal professionals need more than just access to information.”
See also: OpenAI Astra: Delay due to Hugging Face attack
This statement came in support of OpenAI on the day OpenAI began building its own legal search index. OpenAI unveiled Astra for Law on September 17. It’s GPT-6 Astra, the company’s most powerful model yet, configured with legal search, legal analysis and writing guidelines, and controls targeted at law firms. Select firms will first have access through a Trusted Access on ChatGPT and Codex, with the API to follow later as gpt-6-astra-law.
The configuration of the model is less important than what lies beneath it. OpenAI has built a legal search index that covers US case law, statutes, regulations, court rules, and administrative decisions. It covers more than 230 million URLs, with sources added daily.
Much of the case law comes from the Free Law Project, the nonprofit organization behind CourtListener. OpenAI states that this collection covers more than 99.9% of published U.S. case law.
The company is pitching it as complementary. The index “complements the licensed content and specialized products that companies rely on from providers like Thomson Reuters,” its announcement said. Thomson Reuters owns Westlaw, sells CoCounsel and has spent the year building its own legal models. Its CTO provided a statement in support of the launch.
OpenAI tested the configuration on 200 US legal research questions from Vals AI's private validation set, Legal Research Bench. At the highest logic effort, Astra for Legal passed the correctness check on 54.0% of the questions. GPT-6 Astra with only web search passed on 38.7%.
See also: OpenAI releases GPT-6 Astra – AGI has arrived
OpenAI presents this as a 40% improvement in relevant terms, which it is. However, that also means that the trained system fails the correctness check on almost half of the questions. On jurisprudence questions, it found 24% more case references and retrieved up to 54% more relevant excerpts from correct opinions.
The test is by OpenAI, as is the side-by-side comparison it published against Claude Fable 5.1, where OpenAI claims the competing model cited a decision that an appeals court had overturned and said it found no relevant case. No independent entity has verified any of the results.
Alongside the model, OpenAI released 26 plugins built by partners, connecting ChatGPT to tools companies already use, including Relativity, Clio, iManage, Intapp and DeepJudge. Thomson Reuters brings HighQ matter context to ChatGPT and previews a CoCounsel Legal connector. Nine community plugins from LegalQuants, LECG and Skills.law come with 47 custom skills. ChatGPT for Word became generally available on the same day.
The vendors announced themselves in formation. Six of them issued statements between 4:00 PM and 4:15 PM New York time: iManage, Relativity, Consilio, Litera, Laurel and Everlaw. Everlaw went further, allowing lawyers to question the substance of the evidence in the case, not just the metadata. Five of the vendors included the same statement from Jason Boehmig, OpenAI’s general manager for the legal industry, word for word.
Read together, these announcements highlight a common theme. Time entry continues in Laurel. Deeper workflows continue in Litera. Relativity keeps the essential analysis within its own platform. Consilio only passes along material that responds to the question asked.
See also: OpenAI Astra: Pause on most powerful AI model due to critical cyber capabilities
Harvey, valued at $15.6 billion this month, will rely on Astra for Legal via API. Legora, which is seeking a $10 billion valuation, will also rely on it. Harvey’s head of applied research, Niko Grupen, praised the accuracy of its referrals.
