Thomson Reuters Built Its Own AI on Chinese Tech to Cut Its Anthropic Bill

Thomson Reuters launched its first in-house AI model on Monday, August 24, and the foundation underneath it comes from China.

The model, called Thomson, was built on top of Alibaba's open-source Qwen model, according to a Thomson Reuters press release distributed by PR Newswire. The company first retrained Qwen with help from Imperial College London into an intermediate version called Snowdon, named after a mountain in Wales, according to The Decoder. Snowdon was then trained further on Thomson Reuters' own content before becoming Thomson.

Why build it instead of just paying OpenAI or Anthropic more? Money.

The Price Tag

Thomson Reuters CTO Joel Hron told Business Insider the company wanted to rely less on expensive frontier models like Anthropic's Claude. CEO Steve Hasker said on an earnings call this month that the company spent about $40 million training Thomson, according to The Logic. That $40 million covers more than two years of staff and computing costs, not a single training run.

The number that got repeated everywhere, $450,000, is actually just the cost of the final training run for the current version, according to The Decoder and SiliconANGLE, both of which flagged that the headline figure understates the real investment. Add in two-plus years of engineering time, hundreds of subject-matter experts, and decades of proprietary content from Westlaw, Practical Law, Checkpoint and Reuters, and the true cost is considerably higher than the eye-catching $450,000 figure Thomson Reuters emphasized in its own materials.

What It Actually Does

Thomson isn't replacing Claude. It's being deployed first in Tabular Analysis, a document-review feature inside Thomson Reuters' CoCounsel Legal AI assistant, according to SiliconANGLE. CoCounsel will keep using multiple models, leaning on Thomson where its legal specialization helps and on outside frontier models for everything else.

Hron was blunt about the scope. "Thomson needs to set the frontier of intelligence for legal," he told SiliconANGLE. "That's a different job than what I think a lot of the frontier labs are doing." Thomson Reuters expanded its partnership with Anthropic in May for CoCounsel, and Hron told Business Insider that work still relies mostly on Claude.

The Benchmarks Don't Show a Knockout

Thomson Reuters' own numbers, reported by The Decoder, show Thomson trailing competing models on the Stanford LegalBench test, and sitting just behind rivals on the Harvey Legal Agent Benchmark. It falls off sharply on reasoning and coding tasks.

In the company's internal Deep Research benchmark using only open web access, Thomson scored 0.53 on factual accuracy versus 0.65 for a leading competing model. Thomson only pulled ahead, 0.83 to 0.82, when given access to Thomson Reuters' own proprietary content. Evaluation lead Andrew Bean acknowledged to The Decoder that with web access alone, Thomson is "within the scope of the other models, but certainly not the leader yet." The competing model improved almost as much as Thomson did when given the same proprietary content, suggesting the data access matters as much as the specialized training.

Investors Weren't Thrilled

Thomson Reuters' AI strategy has been rattling its stock for months. The Logic reported that shares fell sharply in February after Anthropic launched its own legal tech tool, spooking investors who worried general-purpose AI would eat into Thomson Reuters' niche legal software business. The company responded by deepening its work with Claude.

More recently, the stock fell notably on the day Thomson Reuters disclosed the $40 million training spend during this month's earnings call, according to The Logic. Analysts wrote in research notes that the AI investment will pressure margins that were already tighter than investors expected.

The China Angle

Building an American legal-AI product on top of a Chinese company's open-source model isn't without controversy. Anthropic has publicly accused Chinese AI labs of illicitly "distilling" its models' outputs, and has pushed Washington for restrictions, according to Business Insider. Sen. Tom Cotton has separately raised security concerns about U.S. companies like Airbnb and Cursor building on Chinese open-source models, warning about potential backdoors.

Hron pushed back on the idea that Thomson Reuters is locked into Qwen specifically. He told Business Insider there's "nothing that necessarily ties us to Qwen," and told The Decoder the team has already swapped out the open-source starting point "close to a half dozen times." The company says a secondary alignment process, the Snowdon step done with Imperial College, screened the model for safety, ethics and what Thomson Reuters calls political neutrality.

None of the sources reviewed here show any U.S. government inquiry or restriction currently targeting Thomson Reuters' use of Qwen. Cotton's concerns, as reported, were directed at other companies. Whether that changes now that a Nasdaq-listed company with deep ties to Westlaw and Reuters news content is running a Chinese-derived model in its legal products is an open question nobody in these reports has answered yet.