OpenAI launches Astra for Law: its bet on the legal world starts with a dedicated archive
On 17 September 2026 OpenAI announced Astra for Law, opening a strategy built around vertical editions. The product is not a new model: it pairs GPT-6 Astra with a search index and instructions written for the legal sector. The archive covers more than 230 million URLs of US statutes, regulations and case law, and incorporates data from CourtListener, the non-profit collection run by Free Law Project. The company claims coverage of more than 99.9% of published US precedential case law. Initial access is reserved for a selection of law firms through the Trusted Access programme inside ChatGPT and Codex; an API labelled gpt-6-astra-law is announced as coming later, with no date attached. OpenAI has said nothing about pricing, or about availability in Europe.
On performance, OpenAI released the results of a test on 200 questions drawn from the private validation set of the Legal Research Bench run by Vals AI. At the highest reasoning level, Astra for Law passed the correctness check in 54.0% of cases, against the 38.7% recorded by GPT-6 Astra with web search alone. The company also reports a 24% improvement in finding the relevant cases. The tests were run by OpenAI: according to The Next Web, no independent audit by Vals AI of the published figures appears to exist. Joel Hron, CTO of Thomson Reuters, who announced a preview of a CoCounsel connector, said — in the words reported by SiliconANGLE — that legal professionals need far more than simple access to information; John Savva, of the firm Sullivan & Cromwell, among the first users, spoke, according to the quotation carried by Artificial Lawyer, of an impressive depth of research and a feel for authoritative sources. Both sentences reach us through secondary sources: the official announcement page was not readable by the verification tools, which received a 403 error, so there was no way to compare them with the original.
The move places the company in a field already held by Google, which launched Gemini Enterprise for Legal in August, and by firms such as Harvey and Legora. The launch listed 26 partner plugins, from the likes of Clio and DeepJudge, although trade-press accounts diverge on the total count of tools extended by the community. On confidentiality, the company offers Zero Data Retention on the API for eligible firms and disables human review by default for data handled through ChatGPT Enterprise. The whole apparatus remains confined to US case law.
Choosing to build a dedicated index says that, for OpenAI, generic web search was not enough; how far that index goes is for now measured only by the party selling it. Even in the vendor's own figures, nearly one answer in two fails the correctness check.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- Web search on AI news from the past seven days, in Italian and English. The official announcement on openai.com would not open: HTTP 403 on two attempts, 503 with an alternative reader. I read three independent sources in full — SiliconANGLE (17/09), The Next Web (18/09), Artificial Lawyer (18/09) — plus the passages of the announcement returned verbatim by search engines and the Vals AI methodology page for the Legal Research Bench, which confirms a private validation set distinct from the test set used for the public leaderboards. The figures (230 million URLs, 54.0% against 38.7%, +24%, 26 plugins, Zero Data Retention) match across the three outlets; The Next Web explicitly labels the 54% as self-reported and not independently audited. I dropped the story about the US intelligence report: anonymous sources in a single outlet, and no confirmation or denial from the Department of Defense.
- Incertezze
- The official page was not readable (403): the contents of the announcement come from quotations in several independent outlets and from search results, not from the original document. The two statements quoted come as rendered by SiliconANGLE and Artificial Lawyer and could not be checked against the original. The 54.0% is measured by OpenAI on a private validation set, not published by Vals AI: the comparison with other legal products is therefore not verifiable from outside. Sources diverge on the plugin count (26 from partners; SiliconANGLE adds 9 external plugins with 47 custom skills, Artificial Lawyer speaks of 47 community plugins). There is no pricing, no API opening date, no Trusted Access admission criteria and no indication of any coverage of European or Italian law. Some outlets date the launch to 18 September: most likely their own publication date, not confirmed by an official source.
- Perché pubblicarla
- This is OpenAI's first vertical aimed at a regulated profession, and it shifts the competition from the model to the archive: whoever owns the index of legal sources decides what the model can cite. Two things matter here: a non-profit like Free Law Project becomes the infrastructure of a commercial product, and a declared accuracy of barely more than one correct answer in two — measured by the seller — enters a sector where an error carries professional liability and, under the AI Act, transparency obligations. A clean example of a product that offloads verification onto the buyer.