OpenAI on Thursday launched Astra for Law, a legal-industry version of its GPT-6 Astra model that combines the model with a search index of US case law, statutes, regulations and court rules, plus instructions tuned for legal analysis and drafting. The company is selling it as a foundation for law firms and legal technology companies to build on, not a finished product, and early access runs through a Trusted Access program in ChatGPT and Codex before an API opens later.
The core of the launch is the Legal Search Index. It covers more than 230 million URLs of US legal material, with sources added daily, and integrates the Free Law Project’s CourtListener collection, which OpenAI says covers more than 99.9% of published US case law that carries precedential value. The point, per OpenAI’s announcement, is to move a lawyer from the facts of a matter to the specific authorities and passages that support an answer, which the lawyer can then check directly.
The configuration goes beyond search. Custom instructions for legal analysis and writing guide the model in applying research to a client’s facts, developing arguments or deal terms, and identifying weaknesses and uncertainty in its own draft work. OpenAI describes the result as an environment for legal work rather than a chatbot with a law degree, and the framing matters for how firms end up deploying it.
Benchmark numbers
OpenAI tested the configuration on 200 questions from the private validation set of Vals AI’s Legal Research Bench. Astra for Law passed the overall correctness check in 54% of cases, against 38.7% for GPT-6 Astra using web search alone, a 40% relative improvement. On case-law-focused questions, it identified 24% more relevant cases at the highest reasoning effort.
Those numbers deserve context. A 54% pass rate on hard legal research questions is nowhere near a lawyer’s standard, and the benchmark is OpenAI’s own choice. But the comparison being made is not against humans, it is against the same model without the index, and that gap is the argument for the whole configuration approach: retrieval over a curated corpus beats general web search for this kind of work. The delta between 38.7% and 54% is what the index and the instructions bought, and it is the number competitors will have to match.
Who is building on it
The launch ships with 26 partner plugins, including Thomson Reuters, Intapp, Harvey, Legora, DeepJudge and iManage, plus 9 community plugins from legal engineers at LegalQuants, LECG and Skills.law, alongside 47 adaptable skills for practitioners working in ChatGPT. API customers named at launch include Harvey and Legora, two of the best-funded legal AI startups, which means the platform is partly competing with and partly powering the companies that defined the category.
Sullivan & Cromwell is the named early adopter. Partner John Savva told Legal IT Insider that the product presents “a significant opportunity to consolidate two workflows that have, until now, remained largely separate: researching the law and analyzing legal sources.” He added that it brings the profession closer to conducting research and analysis under one hood.
Why the legal market
Legal is an obvious target for a model vendor. The work is text-heavy, billed by the hour, and organized around citation to authoritative sources, which maps directly onto what a retrieval-augmented model does well. Law firms spend enormous sums on research platforms like Westlaw and LexisNexis, and a model that shortcuts that research while citing the same authorities attacks a real line item in firm economics.
It is also a market where a wrong answer carries malpractice exposure, which is why OpenAI keeps emphasizing that outputs come with citable sources the lawyer examines personally rather than bare conclusions. The company’s own materials stress that the lawyer can “examine for herself” the authorities behind any answer, language aimed directly at the trust problem that has slowed AI adoption in professional services.
The competitive picture is crowded. Harvey, Legora, Clio, Relativity and Thomson Reuters all sell legal AI, and most of them already run on top of one frontier model or another. OpenAI’s move is partly defensive: if legal AI companies are going to build on GPT-6 anyway, OpenAI would rather sell them a configured vertical product with an index moat than plain API access. The index is the differentiator, since assembling and maintaining 230 million URLs of current US law is expensive and dull work that startups rarely do well. Thomson Reuters participating as a plugin partner while selling its own research products shows how entangled the ecosystem has become.
What it does not solve
Hallucinated citations remain the failure mode that got lawyers sanctioned in earlier AI adoption waves, and a better index reduces but does not eliminate the risk of a plausible-looking wrong authority. OpenAI’s own benchmark framing, a 54% correctness pass rate, implies the system is wrong nearly half the time on hard questions. Firms deploying it will need review workflows that assume error, not ones that assume competence.
There is also the confidentiality question. OpenAI says selected firms get early access with protections for confidential client work, but large law firms have been slow to move client data into any third-party model environment, and the Trusted Access arrangement will be scrutinized by every general counsel whose work touches the platform. A single mishandled privileged document would set adoption back years.
The launch is the latest sign that model vendors are moving from selling general assistants to selling configured verticals, and legal is the first profession where OpenAI has built the full stack: model, index, instructions and partner integrations. Whether lawyers actually trust it will show up in firm deployments over the next few quarters, not in launch-day benchmarks. The early signals, a top firm signed on and the category leaders signed on as builders, are about as good as a launch day gets.