--- title: "OpenAI Is Now Getting Into Legal: What Astra for Law Ships - Carlos Arias" description: "Astra for Law is real, gated, and narrower than the headline. What actually shipped, and the half of your pipeline it leaves untouched." url: "https://carlosarias.com/blog/law-firm-marketing/openai-is-now-getting-into-legal" --- [Law Firm Marketing](/blog/categories/law-firm-marketing) # OpenAI Is Now Getting Into Legal: What Astra for Law Ships Astra for Law is real, gated, and narrower than the headline. What actually shipped, and the half of your pipeline it leaves untouched. [Carlos Arias](/blog/authors/carlos-arias) · September 20, 2026 · 7 min read ![Brush-and-ink drawing of a gavel resting on an open book, with a small red seal stamp at the left.](/_astro/cover.dz-i_59O_ZQwRLo.webp) *Brush-and-ink drawing of a gavel resting on an open book, with a small red seal stamp at the left. AI-generated illustration by Carlos Arias .* The thing OpenAI actually shipped into law is narrower than the headline. On September 17, 2026 the company released Astra for Law, its GPT-6 Astra model configured for legal work with a dedicated legal search index behind it. Access is gated to selected U.S. firms, Am Law 200 first. It is a research and drafting instrument for lawyers who already know enough to check it. It will not get your firm found. It will not answer your phone at 7pm on a Friday. Both things are true at once. Most of the coverage only handled the first one, so let’s walk the release properly and then talk about what a principal with six lawyers should actually do this quarter. ## What OpenAI actually shipped Start with the boring accuracy, because the secondhand write-ups are already sloppy about it. Astra for Law is not a new base model. It’s a configuration: legal-specific instructions and settings on top, plus a new Legal Search Index covering U.S. case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with sources added daily (SiliconANGLE). The case law under it comes from the nonprofit Free Law Project’s CourtListener database. That is a meaningful engineering choice, and it’s the part of the release I’d point at first. The rest of the package: - 26 partner plugins at launch, from Thomson Reuters, Intapp, Harvey, Legora, DeepJudge and iManage, plus nine plugins built by lawyers and legal engineers and 47 custom skills (Bob Ambrogi’s LawSites). - Access is gated. It appears in the ChatGPT and Codex model picker as GPT-6 Astra Law for firms admitted to Trusted Access. OpenAI names Wachtell Lipton, Latham & Watkins, Sullivan & Cromwell, Ropes & Gray and Cooley among the firms already working with it (OpenAI). - The API is a promise, not a product. A gpt-6-astra-law endpoint is “coming soon” with no published date and no published price (OpenAI Help Center). - Qualifying firms get zero data retention on the API, and their ChatGPT Enterprise usage is excluded from human review. Nothing in that list is aimed at a solo in Boca Raton. That’s not a complaint. It’s a distribution decision, and it tells you exactly who OpenAI thinks the buyer is. ## The number that should decide how you use it OpenAI published a benchmark, which I respect, and the benchmark says something quieter than the press release. OpenAI tested the configuration on 200 U.S. legal research questions from the private validation set of Vals AI’s Legal Research Bench. At the highest reasoning effort, Astra for Law passed the overall correctness check on 54.0% of questions. GPT-6 Astra with ordinary web search passed 38.7% (OpenAI). That’s a 40% relative improvement. On case-law questions the configured system surfaced 24% more reference cases. Read it again. The winner misses close to half. That is still a real jump, and if you do research for a living it matters. But it defines the product category: this is an instrument that only works with a competent lawyer holding the other end. The verification step isn’t a disclaimer bolted on for the lawyers. It’s load-bearing. We have measurements of what happens when firms skip it. Stanford’s RegLab put 202 queries through the incumbent tools. Lexis+ AI produced incorrect information more than 17% of the time. Westlaw’s AI-Assisted Research managed 33%, despite vendor claims that retrieval “eliminates” hallucination (RegLab). Damien Charlotin’s AI Hallucination Cases database logs only decisions where a court explicitly found reliance on fabricated material. It stood at 1,598 cases on June 9, 2026, and 1,668 by July 2. The count moves daily. I’m not a lawyer and none of this is legal advice. As an engineer, though, I’ll say the design conclusion plainly: any legal AI you deploy needs a named human on the verification step, written down, before the first prompt. ## Google got there three weeks earlier On August 25, 2026, Google Cloud launched Gemini Enterprise for Legal, with Cleary Gottlieb, Freshfields, Weil and Williams & Connolly as launch firms, and connectors into Harvey, iManage, NetDocuments, Thomson Reuters and the same CourtListener database OpenAI is indexing. Two frontier labs decided inside a four-week window that law deserves its own SKU. Artificial Lawyer called the real contest the battle for centrality, and that’s the right frame. Whoever becomes the place a lawyer opens in the morning owns the next decade of that firm’s spend. The pattern is not new. Platforms ship a capability, and the market spends a year arguing about the capability while the revenue settles on whoever already owned distribution. The capability is never the scarce thing. Access to the client is. ## Step by step: what a small-firm principal does with this Here is the sequence I’d run, in order, for a firm under twenty lawyers. - Step 1: Decide whether you’re a buyer or a bystander, and say it out loud. If you’re not Am Law 200, you are a bystander on Astra for Law specifically. The correct response is not to sit in a waiting posture until the API prices. It’s to notice that the part of your business a frontier lab will never build is still sitting there unfixed. - Step 2: Write the one-page AI policy before you buy the tool. What a model may draft, what must be verified against a primary source, who signs, and what happens when the check fails. This takes an afternoon. It is the cheapest risk control available to you, and it’s the thing most firms do last. - Step 3: Inventory the AI already in your building. Clio’s 2026 report on solo and small firms found 71% of solo practitioners and 75% of small firms already using AI, while only about a third had grown revenue from it (Clio). Somebody on your staff is pasting matter facts into a consumer chat window right now. Find out who. - Step 4: Put the agentic layer where the money leaks, not where the prestige is. Research is the glamorous surface. Your leak is downstream of the click. - Step 5: Fix the demand side, because nobody else is going to. Covered next. Steps 2 and 3 are unglamorous and they’re the ones that hold. Step 4 is where I do most of my actual work. ## The half of the pipeline nobody at a frontier lab is building for you Astra for Law improves an hour of associate research. It does nothing at all about the inquiry that came in at 6:40pm and went to voicemail. Clio’s 2024 Legal Trends Report had the research agency Lux phone and email 500 U.S. firms with a genuine client inquiry between June 20 and July 5, 2024. Only 40% answered the phone, down from 56% in 2019. Counting callbacks, 52% were reachable at all, against 73% five years earlier (Clio). Flip it. Nearly half of the profession cannot be reached. I pulled that apart in detail in why law firm leads die after the click. None of it is a model problem. It’s an operations problem with a marketing invoice attached. This is where agentic automation in production actually earns its keep, and where I’ve been building since 2022: an inquiry captured and classified in seconds, routed into the case software you already own, answered after hours by a system that knows your practice areas and your conflicts rules, followed up on a cadence somebody can audit, and attributed back to the source that produced it. Nothing about that is a chat widget. It’s plumbing with judgment in it. The other half is getting found for the right matters in the first place. Search behavior is moving toward answer engines. The language models cite is the language your callers already use, which is why I tell firms to mine their own intake transcripts for the phrasing that earns AI citations before they touch a title tag. LocalSEO, the website, intake and the agentic layer have to be one system. Split across three vendors, they cancel each other out, which is most of what I mean when I argue that a firm is better served by one operator than by an agency bench. ## Where I’d refuse to build A client-facing bot on your public site that answers legal questions. No. Not at a 54% pass rate, and not under ABA 7.1 with sixteen hundred fabricated-citation rulings already on the record. I’d also refuse to rip out Westlaw because a benchmark moved. Replacing a research stack is a fifty-thousand-dollar decision dressed up as a software preference. The honest answer for most firms this quarter: wait for the API. Spend the attention on intake. ## If this is the problem The headline is real. OpenAI is now in legal, and Google beat them to it by three weeks. Both built for firms with a hundred times your headcount. That’s fine. The advantage available to a small firm was never going to be frontier research access. It’s that you can change how your firm answers the phone this month, and a 900-lawyer firm cannot. If that’s the problem you’re actually staring at, write me at hi@carlosarias.com and tell me where your pipeline leaks. I’ll tell you what I’d build, and what I’d refuse to. Tags [#Agentic AI](/tags/agentic-ai/) #Legal AI [#Law Firm Marketing](/tags/law-firm-marketing/)[#Intake Automation](/tags/intake-automation/) Share Written by [Carlos Arias](/blog/authors/carlos-arias) Marketing Engineer for law firms. I combine digital marketing, software, data, automation and AI to improve the whole system — from first click to signed case. On this page - What OpenAI actually shipped - The number that should decide how you use it - Google got there three weeks earlier - Step by step: what a small-firm principal does with this - The half of the pipeline nobody at a frontier lab is building for you - Where I’d refuse to build - If this is the problem ## Continue reading [Law Firm Marketing](/blog/categories/law-firm-marketing) · September 17, 2026 [### What Is a Marketing Engineer? Why I Threw Out Four Job Titles and Kept One](/blog/law-firm-marketing/what-is-a-marketing-engineer-why-i-threw/) What is a Marketing Engineer? The person who builds the system that does the marketing. Here is the honest definition and the search data behind it. Carlos Arias · 8 min [Law Firm Marketing](/blog/categories/law-firm-marketing) · September 11, 2026 [### Why Law Firm Leads Die After the Click (And How to Fix Intake)](/blog/law-firm-marketing/law-firm-lead-intake-marketing-automation/) Most law firm leads die after the click, not before it: unanswered calls, after-hours leakage, broken CRM routing and a single follow-up attempt. Here is the intake system that turns inquiries into signed cases. Carlos Arias · 15 min [Law Firm Marketing](/blog/categories/law-firm-marketing) · September 11, 2026 [### How Law Firm Marketing Actually Works: The Systems Guide (2026)](/blog/law-firm-marketing/law-firm-marketing-systems-guide-2026/) Law firm marketing is not posts and ads — it is the system from search to signed case. SEO, local, paid, website, intake and measurement, with the places firms lose cases they already paid for. Carlos Arias · 18 min ## Stay in the loop. One email when it’s worth it — new posts and updates, no spam. Thanks — check your inbox to confirm. Free. Unsubscribe in one click.