{"id":167954,"date":"2026-11-18T11:01:20","date_gmt":"2026-11-18T19:01:20","guid":{"rendered":"https:\/\/xira.com\/p\/2026\/11\/18\/openai-opens-astra-for-law-to-all-chatgpt-enterprise-customers-and-releases-new-benchmark-results\/"},"modified":"2026-11-18T11:01:20","modified_gmt":"2026-11-18T19:01:20","slug":"openai-opens-astra-for-law-to-all-chatgpt-enterprise-customers-and-releases-new-benchmark-results","status":"publish","type":"post","link":"https:\/\/xira.com\/p\/2026\/11\/18\/openai-opens-astra-for-law-to-all-chatgpt-enterprise-customers-and-releases-new-benchmark-results\/","title":{"rendered":"OpenAI Opens Astra for Law to All ChatGPT Enterprise Customers and Releases New Benchmark Results"},"content":{"rendered":"<p>In September, OpenAI joined other frontier models that have been moving into the legal field, including Anthropic and Google, and launched Astra for Law, a version of its latest GPT-6 Astra model configured specifically for legal work. But it limited access to a handful of so-called Trusted Access firms, including Cooley, Ropes &amp; Gray, and [\u2026]<\/p>\n<p>In September, OpenAI joined other frontier models that have been moving into the legal field, including Anthropic and Google, and <a href=\"https:\/\/www.lawnext.com\/2026\/09\/openai-releases-astra-for-law-a-gpt-6-model-configured-for-legal-work.html\" rel=\"nofollow noopener\" target=\"_blank\">launched Astra for Law<\/a>, a version of its latest GPT-6 Astra model configured specifically for legal work. But it limited access to a handful of so-called Trusted Access firms, including Cooley, Ropes &amp; Gray, and Sullivan &amp; Cromwell.<\/p>\n<p>Now, OpenAI has <a href=\"https:\/\/openai.com\/index\/astra-law-6-2\/\" target=\"_blank\" rel=\"noopener noreferrer nofollow\">released an updated version of Astra for Law<\/a> that is a configuration of its GPT-6.2 Astra model and opened access to every ChatGPT Enterprise customer on a request basis.<\/p>\n<p>In addition to moving Astra for Law onto the newest version of Astra, the release enlarges the legal search index that sits beneath it and adds instructions for researching a question, weighing authorities, and writing an answer.<\/p>\n<p>Perhaps more significantly, OpenAI claims Astra for Law achieved the top score on a benchmark of legal agent work developed by Harvey and is a top performer on the Vals AI Legal Research Bench.<\/p>\n<p>When I <a href=\"https:\/\/www.lawnext.com\/2026\/09\/openai-releases-astra-for-law-a-gpt-6-model-configured-for-legal-work.html\" target=\"_blank\" rel=\"noopener noreferrer nofollow\">first wrote about Astra for Law<\/a> in September, OpenAI called it the start of a long-term investment in law. Jason Boehmig, the former Ironclad CEO who leads OpenAI\u2019s legal products, stressed then that it was not a new model but a configuration of one, adding legal-specific instructions, tooling and a search index. This release continued that approach with the newer base model.<\/p>\n<h3>Benchmark Claims<\/h3>\n<p>OpenAI tested Astra for Law on LAB-AA, the <a href=\"https:\/\/artificialanalysis.ai\/models\/astra-for-law-6-2\" target=\"_blank\" rel=\"noopener noreferrer nofollow\">Artificial Analysis<\/a>\u00a0version of Harvey\u2019s Legal Agent Benchmark. It tests models on 120 assignments that Harvey developed across 24 areas of law, from drafting deposition outlines to reviewing commercial leases. Scoring is all or nothing. An answer earns credit only if it meets every requirement of the assignment without a material hallucination.<\/p>\n<p>Against that measure, OpenAI said, Astra for Law earned full credit on 20.8% of assignments. The base GPT-6.2 Astra model (not configured for law) scored 15.0%, and Muse Wonder Full, the highest-scoring model from another developer, scored 13.1%.<\/p>\n<div id=\"attachment_54889\"><a href=\"https:\/\/www.lawnext.com\/wp-content\/uploads\/2026\/11\/astra-for-law-labaa-ranking.webp\" rel=\"nofollow noopener\" target=\"_blank\"><img loading=\"lazy\" decoding=\"async\" aria-describedby=\"caption-attachment-54889\" src=\"https:\/\/www.lawnext.com\/wp-content\/uploads\/2026\/11\/astra-for-law-labaa-ranking-1024x702.webp\" alt=\"\" width=\"1024\" height=\"702\" title=\"\"><\/a><\/p>\n<p id=\"caption-attachment-54889\"><strong>Source: OpenAI.<\/strong><\/p>\n<\/div>\n<p>As the blog <a href=\"https:\/\/www.orcarouter.ai\/blog\/astra-for-law-6-2-legal-intelligence\" rel=\"nofollow noopener\" target=\"_blank\">OrcaRouter<\/a> points out, that gap in performance between Astra for Legal and GPT-6.2 Astra is larger than it suggests. \u201cIt is the difference between finishing one in five legal assignments cleanly and one in seven.\u201d<\/p>\n<p>Also of note is the potential cost difference. Astra for Law reached its score with less than half the output of Muse Wonder Full, about 64,000 tokens per assignment against 139,000.<\/p>\n<p>With regard to hallucinations, Artificial Analysis counted an average of 0.03 material hallucinations per assignment for Astra for Law, tied for the lowest rate in the comparison. Anthropic\u2019s Claude Opus 5.5, which OpenAI said had the best result of any other developer on that measure, averaged 0.59.<\/p>\n<p>On Harvey\u2019s own LAB, which is separate from LAB-AA, OpenAI said the share of assignments meeting every requirement rose from 4.6% with GPT-6 Astra to 14.6% with GPT-6.2 Astra, and to 16.2% with the legal configuration. That suggests that most of the improvement came from the new base model, with the legal configuration adding another 1.6 percentage points.<\/p>\n<p>OpenAI said that on Vals AI\u2019s Legal Research Bench, where practicing lawyers set questions about U.S. federal and state law, Astra for Law earned full credit on 112 of 208 questions, or 53.8%, up from about 46% in September. That makes it a \u201ctop performer\u201d but not a leader on the Vals bench.<\/p>\n<p>(The 46% September baseline that OpenAI reported this week differs from what it reported at the September launch. At that time, it said Astra for Law passed the overall correctness check on 54% of 200 questions.)<\/p>\n<p>\u201cAstra for Law 6.2 is a material improvement over prior Astra models across our benchmarks, Julio Pereyra, head of legal research at Harvey, said in a statement published by OpenAI. \u201cIt writes more comprehensive and better reasoned briefs that are preferred to those of GPT\u20116 Astra nearly 90% of the time by our pairwise judges. Notably, this improvement in depth comes without regression to prior models\u2019 accuracy \u2014 we continue to see Astra class models have some of the lowest hallucinations rates on LAB.\u201d<\/p>\n<h3>A Larger Legal Search Index<\/h3>\n<p>As of its September launch, Astra for Law\u2019s legal search index covered more than 230 million URLs of U.S. case law, statutes, regulations, court rules and administrative decisions, drawing on the CourtListener database from the nonprofit Free Law Project.<\/p>\n<p>OpenAI said it has since grown the index by 34%, adding more than 80 million URLs, including material from the U.K., Australia and Canada. The index now spans more than 310 million URLs and covers more than 99.9% of published U.S. precedential case law, the company said.<\/p>\n<p>Any ChatGPT Enterprise customer can now request access to Astra for Law through the Trusted Access program in the admin console. Customers can also request API access to build it into their own tools.<\/p>\n<h3>Just the Beginning, Again<\/h3>\n<p>Since May, the legal industry has seen the frontier models move decisively into legal. In addition to OpenAI:<\/p>\n<ul>\n<li><a href=\"https:\/\/www.lawnext.com\/2026\/05\/anthropic-goes-all-in-on-legal-releasing-more-than-20-connectors-and-12-practice-area-plugins-for-claude.html\" rel=\"nofollow noopener\" target=\"_blank\">In May, Anthropic launched Claude for Legal<\/a>.<\/li>\n<li><a href=\"https:\/\/cloud.google.com\/blog\/products\/ai-machine-learning\/introducing-gemini-enterprise-for-legal\" rel=\"nofollow noopener\" target=\"_blank\">In August, Google launched Gemini Enterprise for Legal<\/a>.<\/li>\n<li><a href=\"https:\/\/www.lawnext.com\/2026\/10\/microsoft-launches-copilot-for-legal-becoming-the-fourth-ai-giant-in-five-months-to-build-out-a-legal-vertical.html\" rel=\"nofollow noopener\" target=\"_blank\">In October, Microsoft Launched Copilot for Legal<\/a>.<\/li>\n<\/ul>\n<p>Interestingly, this week\u2019s update from OpenAI focused almost entirely on model performance. It said nothing about new vendor and community plugins or new custom skills, which were part of the September launch.<\/p>\n<p>It also made no mention of access to justice. In September, Boehmig called that a \u201cparamount concern,\u201d and said there would be more to come in the future on that subject.<\/p>\n<p>Interestingly, I closed my September post on the launch of Astra for Law with the subhed, \u201cJust the Beginning,\u201d noting the company\u2019s stated commitment to investing in law for the long term.<\/p>\n<p>In its announcement yesterday, that theme came up again. \u201cWe\u2019re only at the beginning,\u201d it said. \u201cWe can\u2019t wait to see where you take the profession.\u201d<\/p>\n","protected":false},"excerpt":{"rendered":"<p>In September, OpenAI joined other frontier models that have been moving into the legal field, including Anthropic and Google, and launched Astra for Law, a version of its latest GPT-6 Astra model configured specifically for legal work. But it limited access to a handful of so-called Trusted Access firms, including Cooley, Ropes &amp; Gray, and [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":167955,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_et_pb_use_builder":"","_et_pb_old_content":"","_et_gb_content_width":"","footnotes":""},"categories":[24],"tags":[],"class_list":["post-167954","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-lawsite"],"_links":{"self":[{"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/posts\/167954","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/comments?post=167954"}],"version-history":[{"count":0,"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/posts\/167954\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/media\/167955"}],"wp:attachment":[{"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/media?parent=167954"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/categories?post=167954"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/xira.com\/p\/wp-json\/wp\/v2\/tags?post=167954"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}