Thomson Reuters announced on August 24, 2026, that it has developed Thomson, its first in-house large language model for professional legal and tax workflows. The company says it began with an open-source foundation and added proprietary content, training methods and expert input. Thomson's first planned production use is in the Tabular Analysis capability within CoCounsel Legal.
Document review will be the first use case
Tabular Analysis is designed for structured review of large volumes of documents. According to Thomson Reuters, Thomson will become the default model for that capability, while administrators will still be able to select alternatives. CoCounsel Legal will remain a multimodel product, applying Thomson where the company believes its domain specialization provides a clear advantage and using third-party models for other tasks.
That distinction means the announcement is not a wholesale replacement of the models already used in CoCounsel. The company says it will continue to draw on different systems for different jobs. Thomson's availability in Tabular Analysis is tied to an upcoming release, and the announcement does not provide a single general-access date covering every customer.
How Thomson was developed
Rather than training a new foundation model from the ground up, the company says it used an open-source starting point and then applied content and expertise associated with assets including Westlaw, Practical Law, Checkpoint and Reuters. Hundreds of subject-matter experts reportedly helped define training goals, produce examples and evaluate outputs. Thomson Reuters also disclosed that less than 10% of its content has been used to train Thomson so far.
SiliconANGLE reported that the project cost about $40 million over two years, including people and computing. The publication also said the company put the cost of the final training run at roughly $450,000. Those figures come from Thomson Reuters; the launch materials do not include an independent cost audit or a detailed public accounting of the computing used.
Performance claims need context
Thomson Reuters says internal evaluations found Thomson broadly competitive with leading models when each system had access only to the web. It says performance moved to roughly equal or slightly better on some tasks when Thomson was connected to company content. The evaluations reportedly examined answer completeness and whether citations supported claims. A more detailed technical report is expected later.
For that reason, the launch claims should not yet be treated as a comprehensive independent benchmark. The company says it has begun giving legal and AI academics direct access for evaluation and intends to expand external testing over the coming weeks and months. It also plans to release a smaller open-weight version of Thomson on Hugging Face for academic and non-commercial use.
Where it fits in the company's AI strategy
Thomson Reuters presents full control of the model as relevant to governance, deployment and data-handling choices. Its strategy is not framed as competing with the largest general-purpose systems in every field. Instead, it emphasizes specialized training for demanding legal and tax work, where authoritative content and professional review may matter as much as general model scale.
The announcement contains two concrete next steps: integration with Tabular Analysis and external examination through the smaller open-weight version. The company also has plans to extend Thomson models across more of its legal and tax portfolio, but it has not published a timetable for those uses. The verified development at this stage is the launch of Thomson Reuters' first proprietary large language model and its initial focus on structured professional document review.
