Thomson Reuters spent $40M to make a $450K training run worth doing

About $450,000 buys a training run ^[5]. It does not buy the reason to do one. Thomson Reuters' own accounting splits those costs: roughly $40 million over two years on people and computing ^[4], with the final run landing at about 1.1 percent of the total ^[17]. The residual averages something like $20 million a year ^[18], and the source material says where it went. Hundreds of subject-matter experts set training objectives, wrote example legal questions and judged responses in blind comparisons ^[15]. Reinforcement learning taught the model to operate Westlaw and Practical Law ^[8]. Westlaw itself runs to more than 40,000 databases and over 150 years of editorial curation ^[9]. The compute was the cheap line.

So the transferable lesson for a publisher sitting on an archive is not that models got cheap to train. It is that the run has stopped being the constraint, and the curated corpus plus the expert payroll are now the whole cost. Thomson Reuters skipped foundation-model work entirely, starting from an open-weight base and layering on its content, its training methods and professional judgment, which it says cut inference cost as well as training cost ^[6].

The part worth studying is the routing. CoCounsel stays multimodel: Thomson is the default in Tabular Analysis with administrators free to select something else ^[2], while third-party frontier models keep the tasks where the domain model has no edge ^[3]. "Thomson needs to set the frontier of intelligence for legal," Joel Hron, global head of AI and TR Labs, told SiliconAngle, calling that a different job from the one the frontier labs are doing ^[7]. A company that just spent two years building a legal model still sends much of the legal work to somebody else's model by default ^[3]. Read that as a scope statement rather than a hedge: the domain model wins where the proprietary corpus and the in-house tools matter.

The quality evidence is in-house. Thomson Reuters says internal tests had the model broadly competitive with leading models when everything was restricted to the web, and roughly equal or slightly better once connected to company content, scored on answer completeness and on whether citations supported their claims ^[10]. Those results have not had extensive independent validation, and the company is now circulating the model to legal experts and academic institutions, planning a smaller open-weight version on Hugging Face under a noncommercial academic licence, and building a portal for outside API keys ^[20].

Two dependencies sit under the roadmap. Hron concedes the question about keeping pace with faster labs and answers it by arguing that improving open models give Thomson Reuters better foundations for later versions ^[14], which makes another company's release schedule a planning input. And only about 10 percent of the information base has been used ^[11], leaving 90 percent ^[19] that will not simply be poured in: the stated next step is converting the most useful content and product activity into better training signals ^[16]. That is more expert time, the line that costs $20 million a year ^[18], not the one that costs $450,000 ^[5].

Meanwhile the sales argument is ownership. Customer data is not used to train the model, control gives the company more authority over deployment and governance, and it is discussing direct model access with large law firms and corporations that might adapt Thomson to their own knowledge and workflows ^[12].