Thomson Reuters Built Its Own AI Model to Loosen Its Grip on Claude

Thomson Reuters spent roughly $40 million over two years to build its own proprietary AI model, Thomson, on Alibaba's open-weight Qwen3.5-397B, and says it now outperforms Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on legal-specific tasks like LegalBench and PrBench Legal Hard, while the final training run cost only about $450,000. The company is simultaneously expanding its partnership with Anthropic, with the next generation of CoCounsel Legal being rebuilt on Claude's Agent SDK, but the in-house model, developed with Imperial College London, is seen as a strategic hedge against vendor lock-in.

Thomson Reuters just built its own AI model on a Chinese open-source foundation, even as it deepens its paid partnership with Anthropic's Claude. Thomson Reuters spent roughly $40 million over two years building a proprietary AI model called Thomson, and the company says it now holds its own against Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on general benchmarks, and beats all three on legal-specific tasks. That's according to a Thomson Reuters Institute blog post and reporting from the-decoder, which detail a project the company started quietly while it was simultaneously expanding its commercial deal with Anthropic. The timing is the story here. In May, Thomson Reuters and Anthropic announced an expanded partnership connecting Claude directly to CoCounsel Legal, the product more than a million legal professionals across 107 countries now use. The next generation of CoCounsel is being rebuilt on Anthropic's Claude Agent SDK, with general availability expected this summer. Thomson Reuters didn't quietly shelve that work to build a competitor. It kept paying Anthropic with one hand while training a replacement with the other. You don't spend two years and $40 million on a side project. You do it because you've decided that renting your core technology from one vendor is a risk. Not one you want to carry indefinitely. The base layer is Alibaba's open-weight Qwen, most recently Qwen3.5-397B. Working with Imperial College London, Thomson Reuters first retrained that Chinese model for safety, ethics, and what it calls political neutrality. The result: an intermediate version internally named Snowdon, after the mountain in Wales. From there the company layered on pre-training with its own proprietary content and post-training with domain experts. Then came reinforcement learning inside its own tool environments, according to the Thomson Reuters Institute. China Publicly Rejects Anthropic's Claim That Alibaba Stole Claude's Data https://startupfortune.com/china-publicly-rejects-anthropics-claim-that-alibaba-stole-claudes-data/ China's assistant foreign minister Liu Bin dismissed US distillation accusations as "misguided and counterproductive" at the World AI Conference in Shanghai. He never addressed the specific numbers behind Anthropic's claim that Alibaba's Qwen lab used 25,000 fake accounts to harvest 28.8 million exchanges from Claude. - China's response to Claude distillation claims https://startupfortune.com/china-publicly-rejects-anthropics-claim-that-alibaba-stole-claudes-data/ - Alibaba stealing Anthropic AI model data https://startupfortune.com/china-publicly-rejects-anthropics-claim-that-alibaba-stole-claudes-data/ That proprietary content is the real asset: Westlaw, Practical Law, Checkpoint, and Reuters news, decades of material no frontier lab can buy. The company says less than 10% of that trove has gone into training the model so far. That's a floor, not a ceiling. Efficiency is the other surprise. The two-year, $40 million figure covers people and compute across the whole project, but the-decoder reports the final training run itself cost only about $450,000, thanks to gains in training efficiency since the project started. Frontier labs routinely burn nine figures on a single flagship run. Thomson Reuters just showed a legal publisher can get competitive results for a fraction of that, by starting from someone else's open weights instead of pretraining from scratch. How Thomson actually performs against Claude and its rivals On Stanford's LegalBench, Thomson scored 0.823, trailing Gemini 3.1 Pro's 0.843 and GPT-5.5's 0.832, but edging out Claude Opus 4.8's 0.818. On PrBench Legal Hard, Thomson led the field outright at 0.352, ahead of GPT-5.5 at 0.333 and Opus 4.8 at 0.315. On instruction-following, Thomson topped every rival at 0.914. Those numbers, first published by Thomson Reuters itself and independently parsed by the legal-tech site LawNext, show a model that isn't the smartest generalist in the room but wins decisively the moment the task touches its own proprietary archive. That's the whole point. Thomson doesn't have to beat Claude at open-ended reasoning. It has to beat Claude at finding the right clause in Westlaw, and on that ground, the home-field advantage is real. Frankly, this should worry every AI lab selling into enterprise verticals with deep proprietary data of their own. Bloomberg has terminal data. Moody's has ratings history. Any company sitting on decades of documents nobody else can license now has a credible, cheap blueprint for cutting its dependence on a frontier lab, using open weights from Alibaba's Qwen or a similar Chinese release as the base and its own archive as the edge. Anthropic isn't losing Thomson Reuters as a customer, at least not yet. The Claude partnership is expanding, not ending, and CoCounsel's next generation still runs on Anthropic's Agent SDK. But Thomson Reuters no longer needs Claude the way it did a year ago. That shift in power is quiet and financial, and it's now proven out in benchmark numbers. Anthropic's sales team should notice. Vendor lock-in used to be a given for anyone buying frontier AI. Thomson Reuters just showed it's optional. Also read: Druckenmiller Dumps Intel, Micron and Broadcom for AMD and Bitcoin Miners https://startupfortune.com/druckenmiller-dumps-intel-micron-and-broadcom-for-amd-and-bitcoin-miners/ • Google's Gemini 3.7 Flash Beats Rivals on Agent Benchmarks at Half the Price https://startupfortune.com/googles-gemini-37-flash-beats-rivals-on-agent-benchmarks-at-half-the-price/ • Taiwan Indicts Nine, Including an Nvidia Employee, Over AI Server Smuggling https://startupfortune.com/taiwan-indicts-nine-including-an-nvidia-employee-over-ai-server-smuggling/ OpenAI Cuts GPT-5.6 Sol API Prices After Holding the Line for Months https://startupfortune.com/openai-cuts-gpt-56-sol-api-prices-after-holding-the-line-for-months/ OpenAI cut API pricing on its flagship GPT-5.6 Sol model by more than 20% starting August 21, 2026, the first price cut to its top-tier model since launch. The move follows earlier discounts to its Terra and Luna models and comes as Anthropic and Chinese labs like DeepSeek and Moonshot AI undercut OpenAI on enterprise API pricing. - OpenAI GPT-5.6 Sol API price cuts announced https://startupfortune.com/openai-cuts-gpt-56-sol-api-prices-after-holding-the-line-for-months/ - when did OpenAI reduce GPT-5.6 Sol pricing https://startupfortune.com/openai-cuts-gpt-56-sol-api-prices-after-holding-the-line-for-months/