Thomson Reuters is making a clear bet on AI ownership. Instead of relying only on outside models from providers such as OpenAI or Anthropic, the company has built Thomson, its own language model for legal work, using Alibaba's Qwen as the foundation.
The project is not just about having another model. It is about controlling the data, the workflow, the cost structure, and the improvements that come from expert review inside Thomson Reuters products.
A $40 million bet on owning the model
Thomson Reuters spent about $40 million on staff and computing power over more than two years, according to the company. That is far higher than the more widely cited $450,000 figure, which applies only to the final training run for the current version.
Even the $40 million figure does not fully capture what went into the model. The larger asset is Thomson Reuters' own content from Westlaw, Practical Law, Checkpoint, and Reuters, plus the work of hundreds of domain experts.
Thomson is built on Alibaba's open Qwen, most recently Qwen3.5-397B, according to the company. Thomson Reuters worked with Imperial College to retrain the Chinese model for safety, ethics, and political neutrality. That intermediate model is called Snowdon, named after the mountain in Wales.
After that step, Thomson Reuters added its own content, used domain experts for post-training, and applied agentic reinforcement learning inside its own tool environments. So far, less than 10 percent of the available content has been used in training.
Where Thomson performs well, and where it does not
The company's claims about Thomson place it among leading models, but the benchmark picture is mixed. On Stanford LegalBench, Thomson scores 0.823 and trails Gemini 3.1 Pro and GPT-5.5. On the Harvey Legal Agent Benchmark, it sits just behind Opus 4.8.
Thomson leads on instruction following and PrBench Legal. It performs much less strongly on reasoning and especially coding. The comparison is also affected by testing method: Thomson uses test-time scaling, while GPT-5.5 is run without a reasoning mode.
The clearest pattern appears in the company's in-house Deep Research benchmark. With web access alone, Thomson scores 0.53 on factual accuracy, while GPT 5.4 reaches 0.65. When Thomson can use company content, it edges past GPT 5.4 by a narrow margin, 0.83 to 0.82.
Evaluation lead Andrew Bean says that with web access, Thomson is "within the scope of the other models, but certainly not the leader yet." Bean also says there is "a big uplift that comes from being able to train on and practice with your own tools."
That distinction matters. Thomson Reuters' advantage does not appear to come only from model training. The source material and tool access are central to the result, and GPT 5.4 also improves sharply when it is given access to that content.
Why not just fine-tune a frontier model?
Thomson Reuters sees three main reasons to build in-house rather than fine-tune an outside frontier model on legal data.
- Economics: Research chief Jonathan Schwartz says standard fine-tuning techniques "tend to have a strong tendency to degrade general capability." The company also wants to avoid dependence on another provider's inference costs and roadmap.
- Data control: Thomson Reuters believes the performance jump comes from training inside tools such as Westlaw. The company does not want to give that kind of access to an outside provider.
- Compounding value: CTO Joel Hron compares the choice to "renting a house versus buying a house." When experts review product updates, that work can become training data for an owned model.
Hron says that with an in-house model, "you are building equity in something that you own for the long-term, and that compounds over time." With third-party models, Thomson Reuters argues that much of that value stays with the provider.
That does not make the same strategy right for every company. The approach works for Thomson Reuters because it has exclusive data holdings, hundreds of full-time domain experts, and workflows where quality can be measured objectively. Companies without those ingredients may mainly be buying ongoing maintenance costs when they build their own model.
The first use case is document review
At launch, Thomson is taking over the Tabular Analysis feature in CoCounsel Legal. That is a high-volume document review use case where a smaller, cheaper model can make economic sense.
CoCounsel Legal will remain multi-model, and administrators can switch models. Thomson is not being positioned as the orchestrator. Instead, it is intended to handle narrower subtasks such as citation checking.
The company says customer data does not go into training. That point is important because Thomson's strategy depends heavily on proprietary material, expert feedback, and controlled product environments.
A smaller version of Thomson is coming to Hugging Face as an open-weight model under a non-commercial license. Thomson Reuters also plans to release a technical report and developer portal. Hron says early, still non-binding talks are underway with law firms about direct licensing.
What the move says about AI strategy
Thomson is not a simple story about an in-house model beating the biggest general models across the board. The benchmark results show a more practical lesson: specialized AI becomes most valuable when it is connected to proprietary content, expert workflows, and measurable tasks.
For Thomson Reuters, owning the model is a way to keep more of that learning inside the company. The model can improve through the same legal and professional information workflows that the company already operates.
Hron sums up the strategic point by saying the next AI advantage "will come from knowing how to orchestrate it, and knowing which intelligence is important enough to own." Thomson is Thomson Reuters' answer to that question: legal AI built around the information and review processes it already controls.