Harvey built an $11 billion legal-software business on top of other companies’ AI models. Now it’s trying to prove it can build one of its own.
On Tuesday, Harvey introduced Harvey Tenet, its first in-house, proprietary model for legal work. Tenet is designed to help Harvey’s software take on more of the tasks typically done by lawyers over hours or days, at a lower cost than the third-party models it relies on.
The move comes as the companies behind the biggest general-purpose models are circling the legal market. Anthropic has been chasing lawyers with plugins for document review and drafting, while OpenAI has hired Ironclad founder Jason Boehmig to lead its push into legal. Google and Meta may not be far behind.
Their sudden interest raises an uncomfortable question for Harvey. What happens when your supplier decides it wants your customers, too? And how long until one of them catches up to Harvey?
Building a bespoke model could give Harvey more control over both its costs and its fate.
Like many startups, Harvey builds on a smorgasbord of models from companies like OpenAI and Anthropic. Every time a lawyer uses one of those models through Harvey, the company has to pay the model provider for the call — a cost that can rack up fast as usage grows.
A capable model of its own could let Harvey route more work through its own engine, reducing the hefty fees it pays to use outside models. That could offer a path to better margins without asking customers to pay more.
Cost was one motivation. Quality was another, said Gabe Pereyra, the former Google DeepMind researcher who left to start Harvey with Winston Weinberg. He noted that Harvey already routes different tasks to different models based on what they’re good at. He argues that Tenet will give customers another option in that mix — one that’s been shaped around the work they actually care about.
To build it, Harvey first needed to create data that could teach a model how lawyers think. So the company hired attorneys, on staff and on contract through companies like Mercor and Snorkel, to dream up mock disputes and case files, then grade the models on how well they reasoned through them.
From there, Harvey used the material to train a version of Kimi K3, a low-cost, open-source model from the Chinese startup Moonshot. Since its July release, the model has whipped the tech world into a frenzy over its power and price.
Tenet is part of a broader rollout the company is calling Harvey 2. Anique Drumright, Harvey’s chief product officer, said it’s also adding a new “Memory” feature that lets users save preferences about how they work, so its agents can carry those instructions across tasks.
Harvey says it will soon release research showing how Tenet stacks up against other models on legal tasks. Its own benchmark results deserve some skepticism, though. Model makers use these tests to measure where their systems fall short, then train them to do better. That’s useful, but over time it’s a bit like taking an exam after helping write the answer key.
Tenet isn’t live in Harvey yet, and the company won’t say when it will be. Pereyra wouldn’t name any of the law firms that might be testing it. But he wasn’t shy about his ambitions.
Eventually, he wants Tenet to become a building block for law firms to train their own models.
Every legal matter teaches a lawyer something. How to negotiate a tricky clause. How a buyer might react to a certain term. Much of that know-how stays trapped in a lawyer’s head or buried in old documents. Pererya thinks a model built inside a firm’s own walls could help unlock it, turning decades of work into blueprints for agents designed to take up those tasks.
Harvey’s bet is that law firms won’t be building those models from scratch. Harvey would give them Tenet as the starting point, then let each firm train its own version on how its lawyers work.
That would push Harvey into a different kind of business. It starts to look less like a software provider and more like a Big Four professional services firm, not just selling technology, but helping clients configure it around what makes their businesses different.
And that could create a neat irony for a company long dismissed as a “ChatGPT wrapper.” If Harvey can pull that off, the wrapper starts to look like the most valuable layer in the stack.
Read the full article here
