OpenAI's Astra for Law Brings GPT-6 to Am Law 200 Firms

OpenAI spent Wednesday convincing Big Law it needs a different kind of GPT-6. Astra for Law, released September 17, wraps the company's flagship Astra model in a legal-specific search index, a bundle of 26 vendor plugins and a program aimed squarely at the Am Law 200, the top 200 U.S. firms by revenue. It is not a new model so much as a new configuration: same weights, different scaffolding, aimed at a profession that bills by the hour and cannot afford a citation that turns out to be fake.

• Astra for Law pairs GPT-6 Astra with a legal search index covering 230+ million U.S. case law, statute, regulation and court-rule documents, sourced from the nonprofit Free Law Project's CourtListener database.
• On Vals AI's Legal Research Bench, a 200-question independent test, Astra for Law scored 54% overall correctness versus 38.7% for plain web search, a roughly 40% relative jump, while surfacing 24% more relevant case citations.
• It launches with 26 vendor plugins (Thomson Reuters, Harvey, Legora, iManage, Intapp, DeepJudge), 9 community plugins and 47 shared custom skills from practicing lawyers.
• Access starts with API customers including Harvey and Legora, plus a new "Trusted Access" program that offers zero data retention to selected Am Law 200 firms; ChatGPT for Word also went generally available the same day.

What actually shipped on September 17?

Three things bundled under one name. First, a legal search tool wired into GPT-6 Astra that queries CourtListener's archive of U.S. caselaw, statutes, regulations, administrative decisions and court rules, rather than relying on the model's training data or a generic web search. Second, a plugin marketplace: 26 tools built by legal vendors, 9 more from independent legal-engineering shops (LegalQuants, LECG, Skills.law), and 47 reusable "skills" that OpenAI says came directly from lawyers who had already been building their own ChatGPT workflows. Third, a distribution plan: immediate API access for Harvey and Legora, then a "Trusted Access" invite program that offers Am Law 200 firms zero data retention and keeps their prompts out of OpenAI's human review pipeline, terms that matter enormously to a general counsel worried about privilege.

RelatedOpenAI Launches GPT-6 Astra, Its First 'Critical'-Risk Model

Jason Boehmig, the Ironclad co-founder OpenAI brought on to lead legal product, and OpenAI engineer Sherwin Wu ran point on the build. The choice of Boehmig is itself a signal: Ironclad sells contract lifecycle software to exactly the corporate legal departments OpenAI now wants to sell into directly.

Does it actually get the law right?

This is the part legal AI has repeatedly failed at in public. Lawyers have been sanctioned for filing briefs with hallucinated case citations since ChatGPT's first year, and every legal AI vendor now gets judged against that history. OpenAI ran Astra for Law against Vals AI's Legal Research Bench, an independent 200-question test of U.S. legal research, and reported 54% overall correctness against a 38.7% baseline for a model doing plain web search, a 40% relative improvement. It also found 24% more relevant case citations and, on some queries, up to 54% more relevant passages than the baseline retrieved. Answers ran roughly twice as long as the unmodified baseline's, which tracks: a model pulling from an actual index of statutes and case text has more to cite than one guessing from memory.

Fifty-four percent is not a number that inspires blind trust, and it shouldn't. It is a real jump over the alternative, not a claim of solved legal research. Read plainly, it says Astra for Law gets roughly half of a structured 200-question legal test right, which is exactly why OpenAI is pairing the launch with zero-data-retention terms and a slow, invite-only rollout to firms rather than a self-serve public switch.

Why sell the picks-and-shovels to the miners?

The stranger part of this launch is who is standing next to OpenAI in the announcement: Harvey and Legora, two legal AI startups that exist because GPT-4 and GPT-5 weren't good enough at law on their own. Harvey has raised money at a multi-billion-dollar valuation building exactly the kind of legal workflow tooling that Astra for Law's plugin layer now formalizes. Handing both of them day-one API access to a legal-tuned Astra looks, on its face, like OpenAI equipping the companies most exposed to being disintermediated by its own base model.

The more likely read: OpenAI needs Harvey and Legora's accumulated workflow knowledge, their firm relationships and their UI more than it needs to compete with them this cycle. A foundation model company selling a bare API to Am Law 200 general counsel is a much colder pitch than Harvey's white-glove legal product built on that same API. For now the incentive holds both companies in place: Harvey and Legora get a materially better underlying model, OpenAI gets distribution into firms it could not reach cold, and the Trusted Access program lets OpenAI go direct only where a firm is big and sophisticated enough to demand it.

RelatedOpenAI Ships Codex Micro, a $230 Keypad for Coding AI

Who is already building on it?

OpenAI named three firms with custom tools already running on the new stack: Sullivan & Cromwell built an agreement analyzer trained on the firm's own playbooks, Ropes & Gray built an M&A diligence system, and Cooley built GO Public, a capital-markets research tool. All three are Am Law 200 firms with the in-house engineering to build custom tooling rather than wait for a vendor, which is precisely who a Trusted Access invite program is designed to court first: not the median firm, but the ones whose adoption other firms will watch.

Our take

Astra for Law reads less like a legal-AI product launch and more like OpenAI buying itself credibility in a profession that has publicly humiliated AI vendors before. The zero-data-retention terms, the invite-only Trusted Access program and the decision to publish an independent benchmark score rather than a marketing claim are all defensive moves aimed at the exact failure mode that has embarrassed every prior legal AI tool: a hallucinated citation in a real filing. Bringing Harvey and Legora in as launch partners instead of treating them as competitors buys OpenAI time and credibility it could not build alone, at the cost of strengthening two companies that are, structurally, renting their business model from OpenAI's own model improvements. That trade works right up until it doesn't, and the moment to watch is whichever comes first: OpenAI deciding it wants Harvey's customers directly, or a firm's malpractice insurer deciding 54% correctness isn't good enough to indemnify.

Original analysis by GenZTech, drawing on OpenAI's official announcement, LawNext's rollout reporting and the Vals AI Legal Research Bench results. Read OpenAI's announcement.