- Harvey raised $550 million at a $15.5 billion valuation, co-led by Lightspeed Venture Partners and new firm Diffusion.
- The legal AI startup is using the cash to build its own models instead of renting them from OpenAI and Anthropic.
- It also bought AI agent security startup Guardrails AI, its fourth acquisition of 2026.
Harvey just did something odd for a company that OpenAI helped fund: it built its first serious in-house model on a Chinese open-weight system instead of leaning further on its backer’s technology. The move landed the same week the company raised $550 million at a $15.5 billion valuation, more than tripling the $5 billion mark it hit fourteen months ago.
The round was co-led by Lightspeed Venture Partners and Diffusion, a new firm founded by longtime Harvey backer and former Coatue Management investor Kris Fredrickson. Sapphire Ventures and Whale Rock Capital Management joined, alongside existing investors including Sequoia, Kleiner Perkins, Andreessen Horowitz, and Goldman Sachs Alternatives.
“All software companies need to turn into AI companies, full stop. Post-training models is going to become a muscle you need to have to compete as a software company,” chief executive Winston Weinberg said.
Building the models it used to rent
Harvey, founded in 2022 by Weinberg, a former securities litigator at O’Melveny & Myers, and Gabriel Pereyra, a former research scientist at Meta and Google DeepMind, has spent four years routing legal work through models from OpenAI, Anthropic, and Google.
In August, it shipped Tenet, its first proprietary model, post-trained on Moonshot AI’s Kimi K3, an open-weight system out of Beijing, with help from Fireworks AI. For example, a firm running a multi-day contract review can now have that work handled by a model Harvey tuned itself, rather than paying per call to a third party.
The choice of base model is the detail worth sitting with. OpenAI has backed Harvey since 2022 and remains an investor. Building a flagship product on a rival’s open-weight model, one released by a Chinese lab, is a direct hedge against that same relationship, and it puts client work covered by legal privilege on a documentation chain that starts in Beijing rather than San Francisco.
Buying teams rather than products
Alongside the raise, Harvey acquired Guardrails AI, a security platform that stress-tests how AI agents behave, marking its fourth acquisition this year after Hexus in January, the Lume AI team in March, and Benchmark in July.
“Guardrails has spent three years building the answer to that question,” Weinberg said, referring to how firms verify what an autonomous agent will actually do before it touches client work. Terms weren’t disclosed, and Weinberg has said the company treats these deals as acquihires regardless of size, prioritising engineering talent over specific products.
Lightspeed’s Sebastian Duesterhoeft, now a Harvey board observer, has argued that in-house legal teams represent one of AI’s largest remaining addressable markets, calling legal services potentially the second-biggest opportunity in AI after coding. Harvey wants to own that market rather than rent the intelligence that runs it.
A crowded field, and a moving target
Harvey now serves more than 3,000 organisations, up from 1,300 in March, with annual recurring revenue above $400 million. Roughly 80% of Am Law 100 firms use the platform, alongside five Fortune 10 companies and corporate clients, including Latham & Watkins and Microsoft’s in-house legal team.
None of that has slowed the competition. Anthropic has released legal plug-ins for Claude, and OpenAI has partnered directly with law firms to customise ChatGPT, turning two of Harvey’s own model suppliers into rivals.
Closer to home, Swedish rival Legora, whichTech Funding News reported is in talks to raise at more than $10 billion, up from $5.6 billion in March, is scaling just as fast in Europe.
The global legal AI software market is projected to grow from $5.21 billion in 2026 to $40.94 billion by 2034, a 29.4% annual growth rate, and most of that growth is still going toward tools that speed up existing legal work rather than replace the billable hour altogether.
Whether owning the model layer actually protects Harvey’s lead, or just adds a new dependency in its place, is the question its rivals will be watching most closely over the next year.
