Thomson's legal model leads GPT-5.5 by 1.9 points

Thomson Reuters says Thomson-1-Large scored 35.2% on PRBench Legal Hard, 1.9 percentage points above GPT-5.5’s 33.3%, before its first CoCounsel deployment in August. The company’s early benchmark disclosure says the model has trained on less than 10% of its proprietary corpus, while acknowledging different reasoning settings across competitors. Operators should file this beside the widening gap between AI adopters and value capturers: curated domain data and expert evaluation may matter more than buying the largest general model.