The next phase of legal tech evolution is underway with major legal AI companies promising to offer their own models, instead of relying solely on tech giants OpenAI and Anthropic.
Such a move could help companies control their computing costs, according to analysts and consultants, but they also would be responsible for investing in security and model maintenance previously outsourced to companies providing base models.
Harvey on Tuesday said it has developed a custom model called Tenet, and Thomson Reuters late last month said it would soon rollout its own model, called Thomson. In both cases, the companies are using open-source AI technology that is available to download and can be customized. Thomson is usingan unnamed open-source foundation, while Harvey’s Tenet reportedly is built on the open-source model Kimi K3.
The announcements represent a fundamental change in how those companies provide their services. Until now companies including Harvey and Thomson Reuters have been dependent on frontier model labs like Anthropic and OpenAI to provide the computing power that they applied to legal tasks.
Many legal technology companies have acted as model routers — switching between Anthropic’s Claude, OpenAI’s GPT, and Google’s Gemini, depending on which one was best suited for a task. If Harvey and Thomson Reuters can direct more queries to their own models instead of paying for answers from Anthropic or OpenAI, they could boost their profitability, said Patrick DiDomenico, a legal technology adviser and former chief innovation officer at Jackson Lewis.
Still, legal AI companies that opt to rely on their own models will have to take on other duties that go along with that choice, including covering the costs of developing and maintaining a model.
“Now you’re responsible for the security too,” said Brenda Leong, director of ZwillGen PLLC’s AI Division. “And that means that a lot of the things that might be covered — whether well or not — by somebody else, is now on your plate.”
Both companies have for the past couple of years sold tech-anxious attorneys on the idea that they can deliver secure, accurate AI tailored to the unique challenges of attorneys. That pitch came as they relied on the frontier labs to develop and maintain large language models.
“Now you’re not depending on somebody else’s very curated, maintained, updated model,” Leong said. “And so you’re not paying the cost of all of that maintenance and curation — but you have to do it yourself.”
The average attorney using tools built on the new models might not see much of a change, DiDomenico said.
“Unless there’s a drastic change in quality based on the models, then the end user shouldn’t notice a difference,” he said.
But, depending on how good the new models are, lawyers could actually see an upgrade in the broader AI systems built on them, said Daniel Linna, a professor at Northwestern University’s Pritzker School of Law.
“This has great potential to make a big difference in how well these systems function,” he said.
The new models come amid a broader experimentation wave where even Anthropic and OpenAI, are figuring out what their models can do, he said.
“There’s a ton of potential here in these approaches that Harvey and Thomson Reuters are taking,” he said.
Bloomberg Law is a legal technology provider.