Thomson Reuters Launches Its Own Legal AI Model
The data giant spent about $40 million over two years and trained the model on less than 10 percent of its own archive of legal content.
Illustration: a quiet law library at dusk, with cool blue light from a machine room spilling through a glass partition at the far end.
Thomson Reuters launched its own large language model, called Thomson, on August 24, 2026, and it goes to work first inside the Tabular Analysis feature of CoCounsel Legal.
At a glance
- Model name: Thomson, announced August 24, 2026 — built on top of an open-weight base model rather than trained from scratch.
- Investment: about $40 million over two years; the final training run itself cost roughly $450,000, according to SiliconANGLE.
- Training data: Westlaw, Practical Law, Checkpoint and Reuters content — less than 10 percent of it used so far.
- First deployment: the Tabular Analysis feature in CoCounsel Legal; administrators can still switch to other models.
- Planned: an open-weight variant on Hugging Face under a noncommercial academic license, plus an API access portal.
Thomson Reuters unveiled its first in-house large language model on August 24, 2026. The model is called Thomson, and the company says it owns and controls the whole thing outright. It ships first inside Tabular Analysis, a feature of the CoCounsel Legal assistant.
Open weights first, specialization second
This was not a from-scratch build. Engineers took an open-weight model as the base, then layered on proprietary content, in-house training techniques, and input from the company's own subject matter experts. Jonathan Schwartz, who runs foundational research, put it plainly to SiliconANGLE: an open-source model left as-is simply would not know as much.
The training material comes from Westlaw, Practical Law, Checkpoint and Reuters journalism. Westlaw alone spans more than 40,000 databases and over 150 years of legal publishing and editorial curation, per the company. Less than 10 percent of that pool has been used so far.
$40 million buys control of the stack
The company puts the two-year bill at roughly $40 million for talent and compute. The final training run came to about $450,000, according to SiliconANGLE. The argument for that spend is sovereignty: Thomson runs at a fraction of the inference cost of typical frontier models and leans on no outside vendor. Thomson Reuters sells the result under a label of its own, Fiduciary-Grade.
Joel Hron frames the method as starting from a strong foundation, specializing it deeply for the work that actually matters, and ending up with intelligence that is cheaper to run and entirely under your own control. The two sources disagree on his title — SiliconANGLE lists Hron as global head of AI and TR Labs, while the company's release calls him CTO.
What the company leaves out
The release claims parity with current frontier models across several tasks and cites academic testers praising citation quality. It publishes no benchmark scores, names no comparison models, and describes no evaluation protocol. None of those performance claims could be verified for this report. Pricing for the planned API has not been announced either.
Two gaps remain. A third outlet's analysis of what the move means for other SaaS vendors was unreachable at deadline (HTTP 403), so that angle is missing here. And on whether CoCounsel still routes to third-party models, the company answers only indirectly: the assistant stays multi-model by design, and administrators can switch away from Thomson.
What happens next
Thomson arrives with the upcoming CoCounsel Legal release for law firms and corporate legal departments. An open-weight variant is planned for Hugging Face under a noncommercial academic license. A portal giving outside developers API access is still being built.
FAQ
What is Thomson Reuters' Thomson AI model?
It is the company's own large language model, built on an open-weight base and specialized with legal content from Westlaw, Practical Law, Checkpoint and Reuters. It runs first in the Tabular Analysis feature of CoCounsel Legal.
How much did Thomson Reuters spend building Thomson?
The company cites about $40 million for talent and compute over two years. The final training run alone cost roughly $450,000, according to SiliconANGLE.
Will Thomson be released as an open-weight model?
An open-weight variant is planned for Hugging Face under a noncommercial academic license. The company has not given a date for that release.
