Thomson Reuters’ Homegrown AI Model Outperforms Frontier Rivals on Legal Tasks

Thomson Reuters has released benchmark results for its proprietary large language model, Thomson, which demonstrates performance comparable to or exceeding major frontier models such as Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 in specialized tasks. Unlike general-purpose models, Thomson was trained using authoritative content from Westlaw, Practical Law, Checkpoint, and Reuters archives, focusing on high-stakes accuracy in legal reasoning, tax, and accounting.

Key advantages of the model include its ability to provide accurate citations and reduce hallucinations, which is critical for legal and tax professionals. By utilizing domain-specific data rather than broad internet scrapes, the model achieves superior results in professional tasks while remaining smaller and more cost-effective than many frontier systems. The development was accelerated by the 2024 acquisition of Safe Sign Technologies and involves rigorous validation by subject-matter experts.

The model is scheduled for public deployment next month within CoCounsel Legal's Tabular Analysis feature, allowing lawyers to review large volumes of structured documents. This move positions Thomson Reuters as a provider of "fiduciary-grade AI," aimed at professionals in regulated fields who require high precision, accountability, and verifiable reasoning.

(Source：Webpronews)