Power Points: Thomson Reuters has launched Thomson, its first proprietary large language model, trained using its vast stores of legal, tax and professional information. The significance is not simply that another AI model has arrived. Thomson Reuters is attempting to turn the Westlaw and Practical Law data moat into an AI moat, reducing its dependence on OpenAI, Anthropic and other foundation-model providers while challenging the increasingly crowded legal AI market.
Thomson Reuters has formally entered the frontier-model business with Thomson, a proprietary large language model built specifically for professional work. This may be a definative move from Australian-born TR CEO Steve Hasker, (above) using his private banking and McKinsey smarts to bloody the noses of the the multitude of legal AI tools confounding and confusing the legal market.
And unlike the usual AI launch involving several billion dollars, an aircraft hangar full of GPUs and extravagant predictions about humanity’s imminent transformation, Thomson Reuters says it spent around $40 million training the model.
The company says Thomson was developed from a strong open-source foundation and then specialised using decades of proprietary content and expertise drawn from Westlaw, Practical Law, Checkpoint and Reuters.
What is also intriguing is that Thomson Reuters says it has so far trained the model using less than 10 percent of its proprietary content. So the move is aggressive, but also opens up the Thomson Reuters legal motherlode. One day.
The company is positioning Thomson as what it calls “Fiduciary-Grade AI”, meaning AI designed for professions where answers need to be verifiable, auditable and rather more dependable than the confidently invented case citation that has already landed more than a few lawyers before irritated judges.
The Numbers Are Interesting
Thomson Reuters says its internal benchmarking places Thomson competitively against leading frontier models.
Its published results show Thomson scoring 0.352 on PrBench Legal Hard, (confusing?) but reportedly demonstrably ahead of the other frontier models tested. Thomson Reuters also reports competitive performance against models including Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro across various legal and professional benchmarks.
Those results should retain the usual benchmark health warning. They are largely Thomson Reuters’ own evaluations and real-world legal performance is considerably messier than a leaderboard.
The company has, however, begun allowing legal and AI academics to test Thomson independently. Professor Jonathan Choi of Washington University School of Law tested it against ChatGPT and Claude using difficult corporate tax questions and reported preferring Thomson’s answers overall, particularly because of its links to supporting treatises.
Professor Samuel Dahan (left) of Queen’s University and Cornell’s Legal AI Lab said its citation quality was generally competitive with leading frontier models.
Why Thomson Reuters Built Its Own Model
It’s all about money and economics, obviously. And elevating brand awareness when Legal AI is exploding in both development and the ‘silly money’ values we recently reported.
Thomson Reuters already uses external foundation models extensively. Its newly released generation of CoCounsel Legal, for example, uses Anthropic’s Claude Agent SDK and remains deliberately multi-model.
But renting intelligence from somebody else carries an obvious problem.
The supplier can raise prices, change the model or, rather inconveniently, decide that the legal market looks sufficiently lucrative to enter itself.
That possibility is no longer theoretical.
Anthropic has moved aggressively into legal workflows and Google has just expanded Gemini Enterprise for Legal, developed with firms including Cleary Gottlieb, Freshfields, Weil Gotshal and Williams & Connolly. Google’s platform will also connect with Thomson Reuters, Harvey, Legora and other legal technology providers.
Thomson therefore gives Thomson Reuters greater control over both its technology and its economics.
Reuters previously reported that CEO Steve Hasker sees the proprietary model becoming an important part of the company’s legal, tax and accounting products. Around 32 percent of Thomson Reuters’ underlying contract value already involved generative AI during the second quarter of 2026.
The Real Weapon Is Not The Model
The model itself may ultimately be less important than the material sitting behind it.
General-purpose AI companies possess enormous computing power and increasingly capable models. What they do not automatically possess is Thomson Reuters’ enormous collection of edited, classified and interconnected legal material accumulated through Westlaw and Practical Law.
That reinforces a distinction LawFuel has previously examined in The Legal AI Split That Will Reshape Every Law Firm’s Tech Stack: the emerging separation between general or operational AI and systems designed for authoritative professional work.
Thomson Reuters is effectively betting that specialised models + proprietary professional data + human expertise can outperform the simpler formula of giving an enormously powerful general model access to legal documents.
Its early testing suggests there may be something to that argument.
Thomson Is Already Going Into CoCounsel
This is not merely an AI laboratory project.
The first deployment of Thomson will be within Tabular Analysis in CoCounsel Legal, where lawyers can review as many as 10,000 documents and ask up to 100 questions across them.
Thomson Reuters says CoCounsel will remain multi-model, using Thomson where its specialist capabilities provide an advantage while continuing to use other leading models elsewhere.
That is probably the more sensible strategy. The legal AI contest is increasingly unlikely to produce one omnipotent model sitting triumphantly above the profession. Different models are proving better at different jobs.
LawFuel has previously examined precisely this pressure on the legal AI market following Anthropic’s move directly into legal workflows and the increasingly competitive position facing highly valued specialist platforms such as Harvey.
Our Take on the TR Legal AI Move
Thomson may represent something more consequential than another entry in the increasingly tedious procession of “revolutionary” AI launches.
It shows that the legal AI battle is shifting from who has access to the best foundation model to who owns the information, workflows and specialist training capable of making those models genuinely useful to professionals.
It’s the old adage: Content is King. And Thomson Reuters have tons of legal content.
It is rather less comfortable news for legal AI companies whose principal technological moat consists of putting a polished interface around somebody else’s model.
LawFuel examined that vulnerability earlier this year when Anthropic’s legal push rattled legal-tech valuations and exposed what we called the “model + wrapper + workflow” problem.
Thomson Reuters has now taken the opposite approach.
It owns the content. It owns the professional platforms. Increasingly, it intends to own the intelligence sitting underneath them too.
For law firms agonizing over the parade of AI systems, that may ultimately prove more important than whichever chatbot happens to be winning this month’s benchmark beauty contest.
