Thomson Reuters has unveiled its first homegrown large language model, and it's coming for Claude's job. The company spent roughly $40 million on the effort, mostly on talent and compute, to build a model called Thomson that will gradually take over some of the tasks Anthropic's Claude currently handles.
Thomson Reuters CTO Joel Hron told Business Insider that Thomson is built on top of Snowdon, a model developed by a team at Imperial College London by adapting Alibaba's Qwen. Thomson Reuters then continues training it with its own proprietary data from Westlaw, Practical Law, Checkpoint, and Reuters — and so far, it's using less than 10% of the total content available.
Hron compared renting external models to renting an apartment: you keep paying someone else, and you don't control the pricing or the roadmap. The company's legal AI assistant, CoCounsel, still primarily runs on Claude for now, but Thomson Reuters plans to hand more and more responsibilities to Thomson, starting with high-volume tasks like document review.