Thomson Reuters bets $40M on owning its AI instead of renting from OpenAI or Anthropic
decoder ·
Thomson Reuters has built its own in-house AI language model called 'Thomson,' based on Alibaba's open-source Qwen, after spending about $40 million on staff and computing over more than two years. The model was trained on the company's proprietary data and with input from its domain experts, and it only outperforms rivals like GPT-5.4 when it has access to Thomson Reuters' exclusive content and tools. The company chose to build its own model instead of fine-tuning frontier models from OpenAI or Anthropic to cut costs, maintain independence, and avoid being locked into a provider's roadmap and inference pricing. Benchmarks show Thomson leads on instruction following and legal benchmarks but trails on reasoning and coding, and its edge over GPT-5.4 is razor-thin even with exclusive data. The company argues that fine-tuning outside models degrades general capability and that every expert review of an in-house model compounds as training data, like buying rather than renting a house. Thomson will first be used for high-volume document review within the CoCounsel Legal product, and a smaller version will be released on Hugging Face under a non-commercial license. The approach works because Thomson Reuters combines exclusive data, hundreds of domain experts, and measurable workflows, which the article notes most companies lack.
decoder