Thomson Reuters

Legal Tech · Artificial Intelligence · Enterprise Software

Launch

Thomson Reuters launches Thomson, a $40M in-house legal AI model

Built on a Chinese open-source model, it undercuts billion-dollar frontier labs by betting specialization beats scale in legal work.

• Thomson Reuters launched Thomson, its first proprietary large language model, built on Alibaba's open-source Qwen 3.5 and trained for roughly $40 million, initially powering the Tabular Analysis feature in CoCounsel Legal.
• Thomson Reuters (Nasdaq/TSX: TRI) is a global content and technology company whose Westlaw, Practical Law and CoCounsel products form the backbone of legal research and AI tools used by law firms and corporate legal departments.
• Thomson was trained on Westlaw, Practical Law, Checkpoint and Reuters content using pretraining, post-training and reinforcement learning; efficiency gains brought the cost of the final training run down to about $450,000.
• Days earlier, Thomson Reuters made generally available a next-generation CoCounsel Legal, a fully agentic assistant built on Anthropic's Claude Agent SDK that adds a Westlaw Brief Builder for drafting first-draft briefs.
• The launch caps a strategy that began with Thomson Reuters' 2024 acquisition of Cambridge AI startup Safe Sign Technologies, which the company said would accelerate development of legal-specific AI capabilities.
• Thomson Reuters says it will keep a multi-model strategy, using Thomson where its domain expertise wins while still deploying OpenAI and Anthropic models elsewhere, with plans to extend Thomson across the legal and tax portfolio and add sovereign AI options.
• The move signals a shift in professional AI: rather than relying solely on general frontier labs, a legacy content company is wagering that narrow, auditable domain models can match frontier performance at a fraction of the cost for high-stakes work.