Executive Summary
Since February, all three frontier model labs have stood up dedicated legal offerings. Anthropic launched Claude for Legal in February and expanded it in May with more than 20 connectors and 12 practice-area plugins. OpenAI hired Ironclad co-founder Jason Boehmig in June to run a legal vertical, signed Willkie in July, and is developing what it calls Codex for Legal. Google released Gemini Enterprise for Legal in preview on 25 August with Cleary Gottlieb, Freshfields, Weil, and Williams & Connolly as launch customers. In the first week of September, both Google’s Gemini 3.8 Flash release and OpenAI’s GPT-6 Astra launch used legal work as a showcase.
None of this is coincidence. Legal is one of the very few industries where the economics, the data, the buyer, and the competitive landscape all line up in favour of a frontier lab selling directly. This report sets out six reasons why, how the three entries differ, and what remains unresolved. The short version for firms: three of the best-resourced technology companies in the world are now competing to build tools for legal work, and that competition is good news for the firms that engage with it.
Disclosure: the author works in technology leadership at a law firm. This analysis reflects publicly reported figures and the author’s own views, not those of any employer or client.
The week this report was published. The pattern did not stop at the product launches. On 2 September Google released Gemini 3.8 Flash and, in the comparison table for the model, scored it against Harvey’s Legal Agent Benchmark alongside the usual academic tests, reporting 10.0 percent against 6.7 percent for Claude Opus 5 on an agentic legal benchmark where every frontier model still scores in single or low double digits. A day later, OpenAI’s GPT-6 Astra launch video used a law firm as one of its everyday demos: at the 1:22 mark a user asks for a licensing agreement template “for the lawyers at my firm to have a look at”, in a montage that also covers ordering takeout and finding a tennis court. When a legal startup’s benchmark is how the labs grade their own flagship models, and drafting for a law firm is presented as a routine consumer task, the vertical is no longer a side project.
1. Legal Is a Large, Text-Native Market That Is Already Investing
The simplest explanation is that the money is there and it is moving. The Thomson Reuters Institute’s 2026 State of the US Legal Market report found law firm technology spending grew 9.7 percent in 2025, roughly seven points above core inflation and the fastest real growth the report has ever recorded. Knowledge management spend grew 10.5 percent. Technology investment is up 39.3 percent since 2021. Firms are investing from a position of strength: average profit growth of 13 percent, worked rates up 7.3 percent, and demand growth at its strongest since the 2008 financial crisis. Am Law 100 rates now average over $1,000 an hour, a reflection of the value clients place on the work.
That combination matters to a lab trying to turn frontier models into durable enterprise revenue. Law firms have budget authority, high willingness to pay per seat, and a work product that is almost entirely language. Contracts, briefs, discovery, research memos, regulatory analysis, and client correspondence are precisely what large language models do well. There is no physical-world or legacy-hardware problem to solve first. The addressable work is text in, text out, and the value of an hour of skilled legal time makes a per-seat licence one of the easiest technology decisions a firm will make.
2. The Startups Proved the Market, and the Labs Want the Margin Back
Harvey and Legora did the hard work of showing that lawyers will pay for AI at scale. Harvey reached roughly $190 million in ARR by January 2026, raised at an $11 billion valuation in March, and reports over 100,000 lawyers across 1,300 organisations including most of the Am Law 100. Legora raised $600 million at a $5.6 billion valuation and is running a national advertising campaign with Jude Law. Total capital into the two companies exceeds $1.5 billion. Those are the last closed rounds; August reporting has Harvey in talks at around $15 billion and Legora seeking $10 billion or more, which would mean both roughly doubled their paper value inside four months.
Those numbers are a signal the labs could not ignore. Harvey was built on OpenAI models and counts the OpenAI Startup Fund and GV among its investors. The labs have watched application-layer companies capture 50x revenue multiples on top of models the labs themselves supply at commodity API prices. The “for Legal” products are, in part, an attempt to move up the stack and keep more of that margin. Artificial Lawyer framed the OpenAI move bluntly: as the labs approach IPOs, they need sustained enterprise revenue growth beyond general-purpose LLM sales, and legal is one of the most lucrative verticals available.
Partners for now. The labs are not yet trying to kill Harvey or Legora outright. Google lists Harvey, Legora, and Solve Intelligence as partners inside Gemini Enterprise for Legal. Anthropic’s launch included partner-contributed plugins from Harvey and Legora, and Thomson Reuters built a bidirectional CoCounsel integration where Claude can call CoCounsel as a tool. The current posture is “platform underneath everyone,” which is how a hyperscaler or model provider typically enters before deciding whether to compete directly. The startups are not fooled by the partnership language. When Anthropic launched Claude for Legal in May, Harvey CEO Winston Weinberg’s response was “we are prepared to compete,” adding that he and co-founder Gabe Pereyra “have said for years that long term we would end up competing with the model companies.” That tells you how the startups read all three entries.
3. Law Firms Are the Ideal Reference Customer for Enterprise AI
Frontier labs need proof that their models can be trusted with the most sensitive, highest-liability work in the economy. Law firms are that proof. If Freshfields, Cleary, Weil, and Willkie will run frontier models against privileged client matters under ethical walls and confidentiality obligations, the labs can walk into any bank, pharma company, or government agency and say the hardest buyers have already said yes.
Every launch leans on this. Google’s Thomas Kurian stated that raw model intelligence is “nowhere near sufficient” for law and emphasised VPC controls, customer-managed encryption keys, private data isolation, traceable citations, and a hard commitment that customer data never trains the models. Anthropic highlighted Freshfields deploying Claude to 5,700 staff across 33 offices with roughly 500 percent usage growth in six weeks. OpenAI’s Willkie announcement centred on ChatGPT Enterprise plus Codex inside the firm’s own development environment. The legal vertical is a governance showcase as much as a revenue line.
4. Big Law Sits at the Centre of the Economy
Law firms sit at the centre of nearly every high-value transaction and dispute. An M&A deal touches the acquirer, the target, the banks, the auditors, and outside counsel on all sides. If the labs’ models become the substrate on which firms draft, review, and negotiate, they gain a foothold in the workflows of every client those firms serve. Google is explicit about this: Gemini Enterprise for Legal launched alongside Gemini Enterprise for Financial Services, with healthcare, life sciences, and other professional services announced as next. Legal is the beachhead for a professional-services strategy, not a standalone bet.
5. The Technology Finally Matched the Job
Two things changed in the past year that made direct entry practical. First, the Model Context Protocol and similar connector standards let a general model plug into iManage, NetDocuments, Relativity, Everlaw, DocuSign, Box, Datasite, Westlaw, and CoCounsel without each lab building bespoke integrations. All three offerings are built around connectors and practice-area “skills” or plugins, which is a very different product from a chat window. Second, agentic and computer-use capabilities let the models actually execute multi-step legal work: redlining against a playbook, screening regulations, fulfilling DSARs, redacting for motions to seal. OpenAI’s “Codex for Legal” plan is the clearest example, with Codex able to operate desktop applications directly. Once the model can do the task rather than just answer questions about it, the value proposition to a firm changes from research assistant to labour.
6. The Buyer Is Finally Ready
The demand side has shifted too. The Thomson Reuters report found firms with a formal AI strategy are 3.9 times more likely to see critical benefits, and Anthropic’s messaging (“firms that move are pulling ahead fast”) speaks directly to that finding. Legal professionals became the most engaged Claude Cowork users of any knowledge-work function within weeks of the February release, with more than three times the usage of the next function. Managing partners are now competing for clients on AI capability, and the launch quotes from Cleary, Freshfields, Weil, and Williams & Connolly all frame AI as a client-service differentiator rather than a cost-cutting tool.
There is also good judgement in the timing. Clients are asking for efficiency, and this publication’s analysis of general counsel adoption found 87 percent of corporate legal teams now use generative AI, up from 20 percent in 2023. In-house teams and alternative providers are more capable than ever, and Thomson Reuters notes demand growth may moderate by mid-2026 after a strong multi-year run. Firms investing now are positioning themselves to keep serving those clients well whatever the cycle does. That is a sophisticated, well-funded buyer, which is exactly what a lab launching a vertical wants to see.
What Is Different About Each Entrant
The entries are not identical, and the differences hint at where each expects to win.
Anthropic is playing the open-ecosystem card hardest: the widest connector list, open-source plugin contributions from Harvey and Legora, access-to-justice partnerships with Free Law Project and the Justice Technology Association, and a position as the “legal AI fabric” everyone else builds on. Its strength is a model that legal vendors were already adopting and a Cowork product that lawyers gravitated to on their own.
Google is selling infrastructure and governance. Gemini Enterprise for Legal is a hyperscaler product with a control plane for legal IT and risk teams, systems-integrator partners (Accenture, Deloitte, KPMG, Factor), and pre-built partner agents. It is the natural choice for firms already on Google Workspace, and it pointedly includes Microsoft 365 connectors to reach the rest.
OpenAI is betting on distribution and builders. ChatGPT already has the largest installed base of individual lawyers, and the Willkie deal pairs ChatGPT Enterprise with Codex so the firm can build its own tools. Hiring the Ironclad founder signals a contract-lifecycle and in-house-legal orientation, and the forward-deployed engineering model borrowed from Palantir is designed to win the large enterprise accounts that need hands-on implementation.
Microsoft deserves a mention even though it is not a frontier lab. It already sits inside every firm’s Office suite, and its Legal Agent is cited in coverage of OpenAI’s launch as a direct competitor to Codex for Legal. Any firm evaluating the three labs is also, whether it plans to or not, evaluating Microsoft — and the labs’ decision to ship Microsoft 365 connectors acknowledges as much.
Risks and Open Questions
The bull case is strong, but the analysis is incomplete without the caveats.
Hallucination and sanctions risk remains the sector’s defining liability. Fortune’s coverage of the Anthropic launch ran under a headline noting hallucinations are still showing up in filings, and the Mata v. Avianca line of sanctions cases continues to grow. A lab that ships a legal product owns that reputational exposure in a way an API provider does not.
The billable hour is an unresolved conflict. Hourly billing still dominates law firm revenue: 72 percent of firms offer alternative fee arrangements, but fewer than one matter in four is actually billed that way, and Thomson Reuters reports the AFA share of revenue has plateaued. Tools that compress hours change the economics of a matter, and the firms that pair adoption with deliberate pricing will be the ones that benefit; the mechanism, and the case for moving early rather than late, is set out in this publication’s companion analysis. Thomson Reuters makes the same point from the other direction: the firms best positioned are those that tie AI investment to measurable client value.
Commitment is the unknown. Artificial Lawyer laid out three scenarios: the labs go all-in and legal tech consolidates; the labs dabble and incumbents survive with slower growth; or the labs lose interest. The decisive variable is not model capability but how much sustained investment in legal-specific product, sales, and support each lab is willing to fund. Vertical software is a services-heavy business, and the labs have historically preferred horizontal scale.
Incumbents have real moats. Thomson Reuters, LexisNexis, iManage, and Relativity own the systems of record and the authoritative content. Harvey and Legora own adoption playbooks and firm relationships. The labs’ current partnership posture may be a recognition that displacing these is harder than it looks.
What This Analysis Does Not Establish
Vendor figures are self-reported. Harvey’s and Legora’s ARR, user counts and customer numbers come from the companies themselves. ARR is not audited revenue, and the August valuation talks are reported as in progress and may not close at the figures cited.
Usage is not outcome. Freshfields’ 500 percent usage growth and legal’s lead among Cowork functions measure engagement, not the quality or economic value of the work produced. No firm has published outcome data.
The labs’ intentions are inferred. The margin-recapture and reference-customer arguments are this report’s reading of observable behaviour. None of the three labs has stated its legal strategy in those terms.
Bottom line. The surge is the convergence of a rich, text-native, already-spending market; a proven startup playbook the labs want to internalise; the need for high-trust reference customers ahead of IPOs; a leverage point into the broader professional-services economy; connector and agent technology that finally fits the work; and a buyer that is both flush and nervous. For a law firm, the practical implication is that the choice is no longer whether to adopt frontier AI but which platform’s governance model, connector ecosystem, and pricing best fit the firm’s existing stack. Firms have every reason to welcome the attention. Three labs competing for the same customers means better products, sharper pricing, and more leverage for the firm at the negotiating table. The labs are betting that decision gets made in the next 18 months, and they are all showing up at once so they are in the room when it happens.
Sources
Artificial Lawyer. Anthropic Moves Into Legal Tech, 2 February 2026; Claude For Legal Launches, May Reshape the Legal Tech World, 12 May 2026 (including the Weinberg response); OpenAI Plans ‘Codex For Legal’, 18 May 2026; Ironclad Founder Jason Boehmig Joins OpenAI, 1 June 2026; US Law Firm Willkie Partners With OpenAI, 22 July 2026. https://www.artificiallawyer.com/2026/05/12/claude-for-legal-launches-may-reshape-the-legal-tech-world/
ABA Journal. Anthropic launches Claude for Legal, May 2026 — 20 connectors, 12 plugins, Free Law Project and JTA partnerships. https://www.abajournal.com/news/article/anthropic-launches-claude-for-legal-giving-lawyers-20-new-program-integrations-and-12-practice-area-plugins
PYMNTS. Anthropic Scales Legal AI as Lawyers Become Cowork’s Top Users, May 2026. https://www.pymnts.com/artificial-intelligence-2/2026/anthropic-scales-legal-ai-as-lawyers-become-coworks-top-users/
Thomson Reuters. Thomson Reuters and Anthropic Expand Partnership to Connect Claude with CoCounsel Legal, 12 May 2026. https://www.thomsonreuters.com/en/press-releases/2026/may/thomson-reuters-and-anthropic-expand-partnership-to-connect-claude-with-cocounsel-legal
Freshfields. Freshfields and Anthropic Team Up to Co-Build AI Legal Workflows, 23 April 2026 — 5,700 staff, 33 offices, ~500 percent usage growth. https://www.freshfields.com/en/our-thinking/news/news-search/2026/04/freshfields-and-anthropic-team-up-to-co-build-ai-legal-workflows-deploying-claude-across-the-firm-globally
Legal IT Insider. OpenAI launches legal vertical with hire of Ironclad co-founder, June 2026. https://legaltechnology.com/openai-launches-legal-vertical-with-hire-of-ironclad-co-founder/
Google Cloud. Introducing Gemini Enterprise for Legal, 25 August 2026. https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise-for-legal
Thomson Reuters Institute. 2026 Report on the State of the US Legal Market and Will the AI bubble burst?, January 2026. https://www.thomsonreuters.com/en-us/posts/legal/state-of-the-us-legal-market-2026/
Harvey and CNBC, 25 March 2026 — $200M raise at $11B, ARR, customers, investors. https://www.cnbc.com/2026/03/25/legal-ai-startup-harvey-raises-200-million-at-11-billion-valuation.html
TechCrunch, Legal AI startup Legora hits $5.6B valuation, 30 April 2026; Fortune, 18 June 2026 (Jude Law campaign). https://techcrunch.com/2026/04/30/legal-ai-startup-legora-hits-5-6-valuation-and-its-battle-with-harvey-just-got-hotter/
Google DeepMind. Gemini 3.8 Flash model page and launch post by Tulsee Doshi and Raluca Ada Popa, 2 September 2026, including the Harvey Legal Agent Benchmark comparison. https://deepmind.google/models/gemini/flash/; The Register, With Gemini 3.8 Flash, Google reminds everyone it’s still in the race, 2 September 2026. https://www.theregister.com/ai-and-ml/2026/09/02/with-gemini-38-flash-google-reminds-everyone-its-still-in-the-race/5294049
OpenAI. Introducing GPT-6 Astra launch video, 3 September 2026, law-firm demo at 1:22. https://www.youtube.com/watch?v=1QNsdr-Qx_I&t=82s; CNBC, OpenAI announces rollout of GPT-6 Astra model, 3 September 2026. https://www.cnbc.com/2026/09/03/open-ai-astra-gpt-6-cyber.html
Financial Times reporting, August 2026, on Harvey and Legora valuation talks; Fortune, 12 May 2026, on hallucinations in filings; BigHand and Thomson Reuters Institute pricing research on alternative fee arrangements.
