Thomson Reuters built its own model for $40M — data is the asset

Thomson Reuters spent $40 million and came out the other side owning a frontier-class model instead of renting one. The company announced Thomson on August 24 — its first proprietary LLM, trained from an open-source foundation on decades of Westlaw, Practical Law, Checkpoint, and Reuters content. You are almost certainly not going to spend $40 million on a model. That's fine. The transferable part of this story is what made the spend possible, and you may already have a smaller version of it.

What actually happened

Thomson Reuters did not train from scratch. It started from a strong open-source base, then applied mid-training and post-training on proprietary content, with hundreds of in-house subject matter experts shaping the training objectives and the evaluations. The $40 million covers talent and compute — a fraction of what a from-zero frontier run costs.

The claims, which are the company's own and worth reading as such: across legal and general benchmarks, Thomson Reuters says the model performs competitively with the strongest frontier models on the market, including Claude Opus 4.8, and ahead of GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro. CTO Joel Hron's framing in the release is the honest thesis — build intelligence that is "far more efficient and entirely under your control."

Two details make the strategy legible. First, the model has been trained on less than 10% of the company's proprietary content so far — the moat is nowhere near spent. Second, first deployment is narrow: Tabular Analysis inside CoCounsel Legal, with rollout across the legal and tax portfolio to follow. They are also releasing a small open-weight version on Hugging Face for academic and non-commercial use. SiliconANGLE has the outside read.

Why owning your data matters for your business

The asset was never the model. It was fifty years of structured, edited, citation-linked legal content that nobody else has. The model is what turned that archive into a product. Ask the same question about your business: what do you have that a competitor cannot buy? Service tickets with resolutions attached. Ten years of quotes and which ones closed. Install notes, failure modes, part substitutions. Most small companies have this and store it as exhaust.

Structure now, model later. You do not need a training run to benefit. You need that data extractable, consistently labeled, and out of the vendor's proprietary format — because every option downstream, from retrieval to fine-tuning to just answering a question fast, depends on being able to get at it. The companies that will spend on models in three years are the ones cleaning up their data this year.

Narrow deployment is the tell. A $40 million model shipped into one feature first. That's discipline, and it's the same discipline that makes small-company AI projects work: pick the workflow where the answer is checkable, ship it, measure it, then widen. The projects that die are the ones that launched as a platform.

Sources: Thomson Reuters, SiliconANGLE.