Thomson Reuters trained its own AI model. Then it kept using Anthropic’s anyway.

Thomson Reuters has developed its own AI model called Thomson for legal, tax, and compliance work. This model is trained on the company's proprietary professional content and designed to power features like CoCounsel. Early benchmarks show that Thomson competes with models from OpenAI, Anthropic, and Google across several professional and general-purpose evaluations.

However, Thomson was not built from scratch. Instead, Thomson Reuters started with an existing open-source foundation and spent approximately $40 million on training, which included additional training on decades of proprietary content and expert evaluation. The company tells The New Stack that most of their investment went towards training on proprietary content rather than pre-training a foundation model from scratch.

For companies with years of proprietary data, this approach offers an alternative to relying solely on models from OpenAI, Anthropic, or Google, and spending billions trying to build their own. Thomson's training data comes from Thomson Reuters' own collection, including Westlaw, Practical Law, Checkpoint, and Reuters. Hundreds of subject-matter experts were involved in evaluating the model's outputs and identifying areas where it failed. Currently, Thomson Reuters uses less than 10% of the available content for training.

While Thomson is designed for legal, tax, and compliance professionals, it is not the only model Thomson Reuters uses. The company still employs frontier models in other products, such as CoCounsel Legal, which is built on Anthropic's Claude Agent SDK. This SDK is also being used across industries for various applications. However, Thomson serves a specific purpose: legal research and document analysis at scale.

It is the default model for Tabular Analysis within CoCounsel Legal, which can handle up to 10,000 documents and answer up to 100 questions about them. Customers will not be purchasing direct access to Thomson. Instead, it powers specific capabilities within Thomson Reuters' own products. The company is exploring ways to commercialize Thomson in the future.

Building Thomson did not eliminate the problem of uncertainty in legal work. An AI model, including Thomson, is not guaranteed to be error-free. However, owning the training process allows Thomson Reuters to have more control over how the model handles uncertainty. Thomson is specifically trained to admit uncertainty rather than produce confident-sounding answers, which could lead to hallucination or other failure modes that are more problematic in professional work.

Even with this training, Thomson continues to use retrieval, pulling from sources like Westlaw and Practical Law to ensure its responses can be grounded in material that legal professionals can verify. The company employs a combination of training for domain reasoning and retrieval to keep the model anchored to source material. While Thomson is designed to make its reasoning checkable, it cannot guarantee that every piece of an answer can be traced back to a specific passage.

Instead, Thomson Reuters' approach is to ensure that the model's reasoning can be verified where possible, and the final verification is left to the professional using the tool.