Thomson Reuters Launches Thomson-1 to Cut AI Costs

The Canadian data and law group announces Thomson-1, an in-house model built on Alibaba's Qwen lineage, with the goal of reducing its AI costs without forgoing Claude. Developed in collaboration with Imperial College, the tool is launched as voices in the United States express concern over the use of Chinese open-source models, while Anthropic denounces illicit distillation practices.

Anthropic's Accusations and Political Concerns in the United States

Anthropic claims that Chinese labs are illicitly distilling the outputs of its models to train their new chatbots and is calling for the United States to impose restrictions. Senator Tom Cotton has raised alarms about the use of Chinese open-source models by American companies like Airbnb and Cursor, citing security risks, including potential backdoors. Anthropic and Alibaba have not responded to requests for comment. Joel Hron indicates that a joint team has adapted Qwen over several months to develop Snowdon, assuring that it is "ethically and politically balanced" and safe to use. He adds that there is no necessary link connecting the company to Qwen.

AI Costs and the Decision to Invest in Proprietary Assets

For Joel Hron, having an in-house model allows the company to rely on its intellectual property rather than continuously paying third parties. He notes that this decision also aims to leverage Thomson Reuters' expertise in certain sectors, with cost being one of the main reasons, and Thomson-1 being designed to reduce expenses. This strategy is being implemented as large companies face significant AI-related costs, largely due to chatbots like Claude or OpenAI's Codex. Hron likens this approach to buying a home rather than renting.

Planned Uses: Partial Task Management and Document Focus

Thomson-1 will take on certain tasks previously managed by Claude and will focus on areas where the company has strong expertise, starting with document review. Joel Hron emphasizes that this tool is not intended to replace existing collaborations with Anthropic or other labs, and it does not represent a complete alternative to Claude. In May, a collaboration was strengthened with Anthropic around CoCounsel, a legal assistant that, according to Hron, continues to rely primarily on Claude. Thomson-1 was launched on Monday.

Technical Chain: From Alibaba's Qwen to Snowdon and Then Thomson-1

Steve Hasker announces that Thomson-1 has been developed from a realigned version of Alibaba's Qwen3.5. The model is based on Snowdon, which itself originated from the realignment of an open-source Qwen, according to Chief Technology Officer Joel Hron. The open-source nature allows these models to be downloaded for free and adjusted freely, which led to the choice of a Chinese foundation. Several companies are turning to these models due to high AI prices, and a low-cost Chinese AI could represent an asset for Western players; many startups are already building their products on these foundations. Thomson Reuters, the Canadian data and law group, thus positions its initiative within a broader trend where Claude and OpenAI's Codex coexist.