Thomson Reuters has spent more than two years and roughly $40 million building its own legal AI model, and the payoff is a single percentage point. The company’s new in-house model, called Thomson, scores 0.83 on its Deep Research benchmark when it has access to Westlaw and Practical Law content. GPT-5.4, given the same data, manages 0.82. Without that proprietary content, Thomson drops to 0.53 while GPT-5.4 hits 0.65. That’s a sobering margin for a bet this size.
The model was built on Alibaba’s open-source Qwen, most recently the Qwen3.5-397B variant. Thomson Reuters first worked with Imperial College to retrain it for safety, ethics, and political neutrality, producing an intermediate version called Snowdon. Then came pre-training on proprietary content, post-training with hundreds of domain experts, and agentic reinforcement learning inside the company’s own tools. Less than 10 percent of available content has been used so far, which CTO Joel Hron frames as headroom. “The bigger finding is less the individual model and more the model factory we built,” adds research chief Jonathan Schwartz.
On public benchmarks, the results are mixed. Thomson beats rivals on instruction following and PrBench Legal, but trails Gemini 3.1 Pro and GPT-5.5 on Stanford LegalBench and sits behind Opus 4.8 on the Harvey Legal Agent Benchmark. The comparisons are also lopsided: Thomson uses test-time scaling while GPT-5.5 runs without a reasoning mode. Evaluation lead Andrew Bean admits that with web access alone, Thomson is “within the scope of the other models, but certainly not the leader yet.”
The strategic logic goes beyond raw performance. Schwartz argues that fine-tuning frontier models degrades their general capability and locks you into a provider’s pricing and roadmap. Hron frames it as renting versus buying: every expert review during a product update becomes training data for an in-house model, compounding over time. That equity argument only works if you own proprietary data and can measure quality objectively — a profile few companies match. For everyone else, renting from OpenAI or Anthropic may still be the smarter deal.
