The Brief:
Harvey and Thomson Reuters launch their own proprietary AI models, Tenet and Thomson.
The move comes as Anthropic and OpenAI, their own model suppliers, push further into the legal market.
Harvey built an $11bn legal AI business using everyone else’s models. Now it’s building its own.
The company has revealed Harvey Tenet, its first in-house, proprietary model. It’s built to handle the hours-long grunt work lawyers do, cheaper than the third-party models it currently pays for.
Thomson Reuters got there first, unveiling Thomson just weeks earlier. The company says it now ranks among the best in the world on legal and general benchmarks, performing competitively against Claude Opus 4.8 and ahead of GPT-5.5.
Why
The timing isn’t a coincidence. Anthropic has been courting lawyers with drafting and document review plugins. OpenAI poached Ironclad founder Jason Boehmig to run its legal push. Google and Meta probably aren’t far behind.
Harvey and Thomson Reuters’ suppliers are turning into their rivals.
There’s also the money. Every query run through OpenAI or Anthropic costs Harvey and Thomson Reuters a fee. Owning the model means routing work in-house and protecting its margins.
Quality matters too, according to Harvey co-founder Gabe Pereyra. Harvey already sends different tasks to different models depending on strengths. Pereyra says Tenet adds another option, one built specifically around legal work.
Thomson Reuters makes the same bet, leaning on decades of Westlaw, Practical Law and Reuters content to teach its model to reason like a lawyer.
How
Neither firm started from scratch. Thomson runs on an unnamed open-source base. Tenet reportedly runs on Kimi K3, a cheap, open-source model out of Chinese startup Moonshot.
Harvey hired lawyers, staff and contractors through Mercor and Snorkel, to dream up mock disputes and case files, then grade how models reasoned through them.
That work usually happens in lawyers’ spare time. Mercor pays lawyers, retired judges and paralegals US$100 to $200 an hour to invent legal puzzles, deliberately try to trip up the models, and write “golden responses” that teach AI how experienced lawyers actually think.
Thomson Reuters ran a similar playbook, with hundreds of subject matter experts grading outputs and flagging failures. Customer data never trains the model.
What’s next
Thomson launches in August inside Tabular Analysis in CoCounsel Legal, before rolling out across Thomson Reuters’ broader legal and tax portfolio.
Tenet isn’t live yet. Harvey won’t say when it will be, and Pereyra wouldn’t name any firms testing it.
But the ambition is bigger than one model. Pereyra wants Tenet to become the base layer firms use to train their own models on their own precedent and know-how.
Whatever happens next, one thing’s clear: legal AI wants some skin in the model game.
Source: Business Insider, Thomson Reuters, Bloomberg Law