OpenAI said on November 17, 2026, that Astra for Law — the company's legal-specific configuration of its GPT-6.2 Astra model — is being opened to every ChatGPT Enterprise customer through its Trusted Access Program, including API access, after the setup placed first on Harvey's legal agent benchmark.
The move takes a product that had been restricted to a small group of pilot firms since September and makes it available to OpenAI's entire enterprise base, a bet that specialized model configurations — rather than new foundation models — are where domain markets will be won. For more context on this story, see our ongoing artificial intelligence updates.
A configuration, not a new model
Astra for Law is not a separate foundation model. It is GPT-6.2 Astra, OpenAI's newest frontier model, wrapped in legal-specific instructions and paired with a dedicated legal search index. The setup is designed for the core loop of legal work: researching questions, weighing competing authorities and drafting documents.
When OpenAI introduced Astra for Law in September 2026, it said the legal search index covers U.S. case law, statutes, regulations, court rules and administrative decisions, drawing on the Free Law Project's CourtListener collection covering more than 99.9% of published U.S. precedential case law, and complementing licensed content firms already buy from providers such as Thomson Reuters. Harvey and Legora were named as API customers building the configuration into their own products, alongside 26 ecosystem plugins connecting ChatGPT to tools firms already use, such as Relativity and Clio, plus new privacy and governance controls for confidential client work.
First place on Harvey's LAB-AA benchmark
The centerpiece of the expansion announcement is the Harvey LAB-AA benchmark, an Artificial Analysis-run version of Harvey's Legal Agent Benchmark. The test puts models through 120 assignments spanning 24 areas of law, from drafting deposition outlines to reviewing commercial leases to mapping ownership changes in acquisitions. Full credit requires meeting every requirement of an assignment without introducing a material claim unsupported by the source documents.
Astra for Law earned full credit on 20.8% of assignments, the top score. Plain GPT-6.2 Astra managed 15.0%, and the strongest non-OpenAI entrant, Muse Wonder Full, finished at 13.1%. OpenAI said the legal configuration cleared the all-pass bar nearly 40% more often than the runner-up version of its own model — evidence that the instructions and search index, not raw model capability, drove the gain.
The configuration was also markedly cheaper to run. Astra for Law averaged roughly 64,000 output tokens per assignment, 54% below Muse Wonder Full's 139,000 — a gap that matters for legal teams billed by usage.
Hallucination rates: the number lawyers actually watch
Hallucination performance drew particular attention. Astra for Law averaged 0.03 material hallucinations per assignment, tied with GPT-6 for the lowest rate in the comparison and half the 0.06 recorded by regular GPT-6.2 Astra. Claude Opus 5.5, the strongest non-OpenAI performer on the metric, averaged 0.59 — roughly twenty times the Astra for Law figure. OpenAI said the legal configuration logged no material hallucinations on tasks that met the all-pass criteria.
The stakes here are concrete. Courts have sanctioned lawyers for filings citing fabricated cases, and a single invented citation can damage an attorney's standing. For legal buyers, a twentyfold difference in material hallucination rate between competing systems is the kind of number that decides procurement.
Vals AI: the passing rate climbed fast
Independent evidence points the same direction. On the Vals AI Legal Research Bench, where practicing lawyers pose research questions on federal and state law, the share of Astra for Law answers passing every check rose to 53.8% — 112 of 208 questions — from about 46% less than two months earlier. Across all questions, the model captured 90.9% of available points including partial credit. OpenAI attributes the gains to a dedicated legal research team and a search index that has grown 34% since September to more than 310 million URLs.
Harvey's own evaluation team echoed the trend. "Astra for Law 6.2 is a material improvement over prior Astra models across our benchmarks," said Julio Pereyra, head of legal research at Harvey. "It writes more comprehensive and better reasoned briefs that are preferred to those of GPT-6 Astra nearly 90% of the time by our pairwise judges." Pereyra added that the depth gains came without a regression in accuracy, with Astra-class models continuing to post some of the lowest hallucination rates on Harvey's benchmark and its U.S. case law extension.
From pilot to platform
The trajectory is notable for its speed. Astra for Law rolled out in September 2026 to a small group of firms including Cooley, Ropes & Gray and Sullivan & Cromwell. Two months later, the configuration leads the major legal benchmarks, its hallucination rate has halved, and every ChatGPT Enterprise customer can request access.
The strategy is also a signal. Rather than training a bespoke legal model, OpenAI is productizing configuration: domain instructions, a domain search index, governance controls and partner plugins wrapped around one frontier model. If that formula repeats across other regulated industries, the competition in vertical AI will shift from who builds the best model to who builds the best context — and the benchmarks that measure the difference will only grow more consequential.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →