Thomson Reuters Spent $40M to Own Its AI

A 173-year-old legal publisher spent $40 million to build its own AI -- and discovered that the only thing that makes it smart is its own data.

Thomson Reuters launched "Thomson," its first proprietary large language model, on August 24. The model is built on Alibaba's open-weight Qwen architecture, not on a frontier model from OpenAI or Anthropic. The company's reasoning is simple: owning the model is better than renting one, even if the rented one is smarter.

"We don't see owning an AI model that embodies the knowledge and expertise that TR possesses as something that's non-core to what we do," Joel Hron, global head of AI at Thomson Reuters, told SiliconAngle. "AI is a new mechanism for expertise delivery."

The $40 Million Question

The project cost roughly $40 million over two years in people and computing. The final training run of the current version cost only about $450,000 -- a sign that AI infrastructure costs are dropping faster than the headline numbers suggest.

But the real capital is invisible on the balance sheet: decades of content from Westlaw, Practical Law, Checkpoint, and Reuters, plus the working hours of hundreds of domain experts who helped define training objectives, create example legal questions, and judge responses in blind comparisons.

Only about 10% of the company's total information base has been used so far, according to Andrew Bean, a senior research scientist at Thomson Reuters.

The Qwen Foundation

Thomson Reuters began with Alibaba's open-weight Qwen, most recently Qwen3.5-397B. Working with Imperial College London, the company first retrained the Chinese model for safety, ethics, and political neutrality. This intermediate version is called "Snowdon," named after the mountain in Wales.

Then came pre-training on the company's own content, post-training guided by legal professionals, and agentic reinforcement learning inside Thomson Reuters' own tool environments -- Westlaw, Practical Law, and the broader CoCounsel ecosystem.

The company says it has "changed the open source starting point like probably close to a half dozen times already," according to CTO Hron. The bigger finding, adds research head Jonathan Schwartz, is "less the individual model and more the model factory we built."

Benchmarks: The Data Moats Tell the Story

Thomson's benchmark results are revealing -- and not just because of where they lead.

With access only to the web, Thomson scores 0.53 on factual accuracy in the company's Deep Research benchmark, while GPT-5.4 hits 0.65. Thomson is "within the scope of the other models, but certainly not the leader yet," Bean acknowledges.

Connect Thomson to Westlaw content, and the picture flips: Thomson edges past GPT-5.4, 0.83 to 0.82.

On Stanford LegalBench, Thomson scores 0.823, trailing Gemini 3.1 Pro and GPT-5.5. On the Harvey Legal Agent Benchmark it sits just behind Anthropic's Opus 4.8. It leads on instruction following and the PrBench Legal benchmark, but falls off sharply on reasoning and coding.

The data tells a clear story: the model's value is inseparable from the data it was trained on. Give GPT-5.4 the same Westlaw content, and it improves just as sharply. The moat is the content, not the architecture.

CoCounsel: Not a Monoculture

Thomson Reuters is not replacing its existing AI stack. CoCounsel, the company's AI assistant for legal professionals, will remain a multimodel product. Thomson will be the default for some tasks -- starting with Tabular Analysis, a high-volume document review feature -- but administrators can select other models.

This pragmatic approach sets Thomson Reuters apart from companies that go all-in on a single AI strategy. The model is optimized for legal work, not for competing with frontier labs across every domain.

"Thomson needs to set the frontier of intelligence for legal," Hron said. "That's a different job than what a lot of the frontier labs are doing."

Lessons for the Enterprise

The Thomson Reuters case offers a template for other companies with valuable proprietary data. The economics are shifting: building a specialized model on open-weight foundations is becoming cheaper than fine-tuning frontier models, which researchers say "tend to have a strong tendency to degrade general capability."

Thomson Reuters also sees ownership as a governance advantage. Customer data is not used to train the model, and controlling the model gives the company more authority over deployment and future development.

A smaller open-weight version of Thomson will be released on Hugging Face under a noncommercial academic license. The company is also developing a portal where outside developers can request API keys and test the model directly.

The Catch

The benchmark results are not independently validated. Thomson Reuters has begun sharing the model with legal experts and academic institutions for testing, but a promised technical report has not yet been published.

There is also the question of pace. Frontier labs release new models every few months. Can a 173-year-old legal publisher keep up? Hron argues that improvements in open models will give Thomson Reuters stronger foundations for each future version, while the company's investment can remain concentrated on professional work.

The question is whether that concentration is enough. If the value of Thomson is primarily the data behind it, then the model is only as good as the content strategy feeding it. And if GPT-5.6 or Gemini 4.0 can match Thomson's legal performance simply by being connected to the same Westlaw data, the $40 million bet starts to look less like a moat and more like a feature.

But for now, Thomson Reuters has done something few companies have attempted: it built its own AI, on its own terms, and proved that data is the real differentiator -- not the model itself.

Sources

• Thomson Reuters launches proprietary AI model for legal work -- SiliconAngle
• Thomson Reuters bets $40M on owning its AI instead of renting from OpenAI or Anthropic -- The Decoder
• Thomson Reuters goes via China to build its own AI model -- Computing UK
• Thomson Reuters CTO Joel Hron Debuts Thomson Legal AI Model -- AI Magazine