Thomson Reuters trained its own AI model. Then it kept using Anthropic’s anyway.
The New Stack Amanda Caswell ● Covered by 2 sources
Thomson Reuters built its own AI model for legal work, trained on its own content. But it still leans on Anthropic for parts of CoCounsel Legal.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Thomson Reuters has built a house model for legal, tax and compliance work, and it’s not pretending that means the company is going all-in on homegrown AI. The new model, Thomson, was trained on proprietary professional content and is already being used inside products like CoCounsel. Thomson Reuters says it spent about $40 million on the effort, including compute and talent, and that most of the money went into training on decades of its own material and expert evaluation rather than pre-training a foundation model from scratch.
That choice matters because Thomson Reuters has something many companies don’t: years of specialized data sitting behind products lawyers and tax professionals already use. The model draws on content from Westlaw, Practical Law, Checkpoint and Reuters, with hundreds of subject-matter experts helping evaluate outputs and spot failures. So far, the company says it has used less than 10% of the content available to it for training.
But this is not a clean break from the big frontier labs. Thomson Reuters still uses Anthropic’s Claude Agent SDK in CoCounsel Legal, and Thomson itself is being deployed where a domain-tuned model makes the most sense. Right now, that means Tabular Analysis in CoCounsel Legal, which can handle as many as 10,000 documents and answer up to 100 questions about them. Thomson is becoming the default model for that feature, while customers are not buying direct access to it yet.
The company is trying to solve a classic legal-AI problem: confidence can be dangerous. Thomson Reuters says Thomson is trained to flag uncertainty instead of blurting out a polished answer when it doesn’t have one. The model can also retrieve material from Westlaw and Practical Law so its answers stay tied to sources professionals can check, and CoCounsel adds citations, verification and review around that work.
The benchmark story is mixed, which is probably the honest part. Thomson Reuters says Thomson beat Gemini 3.1 Pro, Claude Opus 4.8 and GPT-5.5 in three of seven reported categories, but other models led elsewhere, and the tests were not run the same way. Thomson Reuters also gave the models 53 legal research questions written by its own experts, with Thomson using Westlaw and Practical Law through an internal agentic system while the others searched the web through Brave. That makes the result interesting, not decisive.
My take — AI-written commentary, not fact-checked reporting
This is the sane way to do enterprise AI: rent the generic stuff, own the part where the company actually has an edge. The cute fantasy that every big firm needs its own frontier model has already met the budget. Thomson Reuters is showing the more boring truth — proprietary data beats model cosplay.
Read more about this at: The New Stack
