Thomson Reuters launches in-house LLM Thomson, claiming frontier-level results for $40 million

Thomson Reuters builds its own model

Thomson Reuters has launched Thomson, its first proprietary large language model, developed entirely in-house. According to the company's press release dated August 24, 2026, the model was built by starting from a strong open-source foundation model and investing roughly $40 million across talent and compute — a fraction of the billions that frontier labs have poured into training runs and infrastructure.

The result, the company says, is a model it fully owns, without the heavy inference costs typical of frontier-scale systems. CEO Steve Hasker claims early evaluations put Thomson on par with the latest frontier models across a range of tasks. Those results are the company's own, and broader outside validation is still in progress.

Specialization over scale

The press release positions Thomson as an argument against the industry's default answer of more scale. Instead of raw size, Thomson Reuters leaned on decades of proprietary content — Westlaw, Practical Law, Checkpoint and Reuters — applied through mid-training and post-training, with hundreds of subject matter experts involved from the design of training objectives through to final evaluations.

So far the model has been trained on less than 10% of Thomson Reuters' content, and the company says the next phase is not simply feeding it more data but finding new kinds of specialization.

Reported gains include a meaningful uplift over the base model in instruction following — precisely executing complex, multi-part professional instructions — and an even larger one in reasoning over dense, domain-specific material. Thomson Reuters argues this challenges a common assumption: that a capable general-purpose model only needs access to the right content to perform at expert level. Its early results suggest that proprietary training and human expertise layered onto a strong foundation produce gains that content access alone does not.

"For years, the AI industry has treated scale as the answer: bigger models, more compute, more money," CTO Joel Hron said in the release, arguing Thomson shows another path and "changes the economics of professional AI."

First deployment and open weights

Thomson's first assignment is inside Tabular Analysis in CoCounsel Legal, a feature for reviewing large volumes of structured documents, and it will reach law firms and corporate legal departments in the upcoming release. CoCounsel Legal will continue to run several models by design: Thomson is applied where its advantage is clearest, with other leading models used elsewhere. Thomson Reuters also plans to extend Thomson models across its legal and tax portfolio, with further sovereign deployment options to follow.

A small open-weight version of Thomson is available on Hugging Face for academic and non-commercial use, and the company has begun giving legal and AI academics direct access to evaluate the model. A technical report covers the underlying foundation model.

What early evaluators said

Two academics are quoted in the release. Jonathan H. Choi of Washington University School of Law tested Thomson against ChatGPT and Claude on tough corporate tax questions and said that while all three answered correctly, he preferred Thomson's responses overall, particularly the links to treatises that made answers easier to verify. Samuel Dahan, who directs the Conflict Analytics Lab at Queen's University and the Cornell Legal AI Lab, said his evaluation found Thomson's citation quality generally competitive with leading frontier models, including on Canadian employment-law questions without a Canada-specific setting.

Sovereignty as a selling point

Thomson Reuters is also marketing the model around AI sovereignty: how a model is trained, what behaviors and biases it contains, where it runs, and how client information is protected. The company says customer data is not used to train the model without explicit consent, and it frames the result as meeting a "Fiduciary-Grade" standard aimed at professionals with duties of care and accountability.

Why it matters

One of the world's largest owners of professional content has decided that owning a model beats renting one. If the claims hold up under external scrutiny, Thomson becomes a prominent test of the data-advantage thesis: that a strong open-source base, a deep proprietary corpus and in-house domain experts can produce a specialist model for roughly $40 million that rivals general frontier models where it matters most — suggesting that publishers sitting on high-value corpora may not need to keep paying labs for frontier access indefinitely.

The deployment pattern is equally telling. Thomson Reuters is not ripping out third-party models; it is slotting its own in where it wins. That pragmatic template — build narrow, own the vertical, keep frontier models for everything else — is one other content-heavy companies are likely to study closely.