The GPUs Were the Cheap Part
Thomson Reuters trained a frontier-class model for $450,000 and published enough detail to check the arithmetic. The arithmetic says something different from the headline.
Thomson Reuters trained a frontier-class model for $450,000 and published enough detail to check the arithmetic. The arithmetic says something different from the headline. What happened On 24 August 2026 , Thomson Reuters launched Thomson 1.0 — its first proprietary large language model — after a $40 million programme running two years and covering talent and compute. The figure that will travel is $450,000: the cost of the final training run . It is accurate. It is also about 1% of what the programme cost, and the programme itself excludes the assets that made it possible. The model went live behind a single feature: Tabular Analysis , a high-volume document-review capability inside CoCounsel Legal. Not the assistant. Not the platform. One job. Why it matters Build-versus-buy is normally argued in the abstract by people with no numbers. Here is a company with a proprietary corpus, professional distribution and real liability exposure, publishing what it spent and what it got. For a CEO , the strategic claim is ownership rather than capability. CEO Steve Hasker put it plainly: "The company has always owned the content, the expertise, and the tools professionals rely on every day. Now it owns the model too." When your core product runs on a rented model, part of your gross margin and part of your release schedule belong to your supplier. For a CPO , the lesson is the scoping. Thomson Reuters shipped a risky in-house model as the default for one feature, with a switch in the admin panel . CoCounsel remains multi-model. That is how you deploy something uncertain without betting the product on it. For a founder , the honest read is useful precisely because it is discouraging. $450,000 of compute is within reach of a Series A company. A 175-year archive is not. Under the hood The chain has three layers, and the middle one is genuinely unusual: Qwen 3.5 (Alibaba, open weights) → Snowdon → Thomson 1.0 Snowdon is Imperial College London's model , developed at the FAIR Lab — a joint initiative with Thomson Reuters that emerged as an academic byproduct of the Safe Sign acquisition. An academic artefact sits inside a shipped commercial system, and almost no coverage has mentioned it. The interesting engineering is not the pre-training. It is what they did to stop specialisation from breaking the model. In their own words: "Data mixture matters so much because specialising a model can damage it." They treated "general capability retention as a first-class training objective rather than an acceptable loss." This is the failure mode nobody advertises. Fine-tune hard on legal text and you get a model that knows more law and reasons worse. Their counter was to deliberately train on tax, accounting and world events — domains with explicit chain-of-thought structure — so reasoning patterns came from several places rather than legal vocabulary from one. Then the part that actually cost money: - Partner-level lawyers spent months constructing evaluation rubrics for