Overview
- Thomson Reuters publicly launched Thomson, also called Thomson‑1, on Monday and has put the model into CoCounsel Legal to power the Tabular Analysis feature that reviews large document sets.
- The company built Thomson by adapting an open‑weight foundation—Snowdon, itself realigned from Alibaba’s Qwen—after acquiring Safe Sign and working with academic and internal teams to retrain the model.
- Thomson Reuters says it spent about US$40 million on training, used less than 10% of its Westlaw, Practical Law, Checkpoint and Reuters content so far, and involved hundreds of subject experts in evaluation.
- Early internal and published benchmarks claim Thomson competes with Claude, GPT‑5.5 and Gemini on some tasks, but Thomson Reuters acknowledges methodological differences in the tests and continues to use Anthropic’s Claude for other product capabilities.
- The model is designed to flag uncertainty and use retrieval from Thomson Reuters sources to improve verifiability, but the company warns not every answer can be traced line‑by‑line and final verification remains the professional’s responsibility.