Thomson Reuters Now Owns the Model, Not Just the Content

What Happened

On August 24, Thomson Reuters launched Thomson, the first large language model the company has built and fully owns. It is not a from-scratch foundation model. The team started with an open-weight base — most recently Qwen 3.5 — and trained it into a legal specialist using the company's own archive: Westlaw, Practical Law, Checkpoint and Reuters. The formal name is Thomson 1.0, and the company describes it as the first in a family of models that will carry the same name.

The numbers are the part that made people sit up. Thomson Reuters says it spent roughly $40 million over two years on talent and compute across the whole project — and that the final training run for the version shipping this week cost about $450,000. For comparison, the frontier labs discuss training budgets in the billions. "This $450,000 number I think is indicative of what we were able to do with our data and content on top of world-class open source models," said Joel Hron, the company's chief technology officer and global head of AI.

Thomson's first home is CoCounsel Legal, where it will power Tabular Analysis — the high-volume, structured document review workflow where a purpose-built model has the clearest edge. It becomes the default model for that feature, though administrators can switch it out. CoCounsel itself stays multi-model: Thomson where it wins, third-party frontier models everywhere else. Hron expects that split to shift over time, with Thomson taking "a bigger and bigger share of the tokens."

The training pipeline had four stages, according to Jonathan Schwarz, who heads foundational research at the company. First, realign the base model's values and behavior. Second, continue pretraining exclusively on proprietary content. Third, targeted post-training shaped by subject-matter experts. Fourth — and this is the unusual one — teach the model to actually operate the company's research tools, so it can run a Westlaw query rather than recall a Westlaw answer.

Why It Matters

For three years the assumption in enterprise AI has been that you rent intelligence. You buy an API, you wrap your data around it with retrieval, and you accept that the model underneath belongs to someone else. Thomson is an argument that a company sitting on a genuinely rare corpus can do something different — and that the price of doing so has collapsed far enough to make it a normal capital decision rather than a moonshot.

Westlaw is the reason this works. The platform spans more than 40,000 individual databases and over 150 years of legal publishing and editorial curation — headnotes, key numbers, citator signals, all of it built by human editors over generations. That is not scraped web text. It is structured, adjudicated, continuously maintained knowledge, and it is the kind of asset that does not exist twice. Thomson Reuters has been quietly assembling it since long before anyone thought of it as training data.

The economics follow from the corpus. Because the model is small and specialized rather than large and general, inference costs less. Because the training run is a fine-tune of an open-weight base rather than a ground-up pretrain, the capital outlay is a rounding error against what OpenAI or Anthropic spend. And because Thomson Reuters swapped its open-source starting point roughly half a dozen times over the project's life as better bases appeared, the company effectively gets to ride the open-weight improvement curve for free. "The bigger finding here is less the individual model and more the model factory we built," Schwarz said.

There is a governance argument underneath the cost argument, and it may matter more to buyers. Thomson Reuters says no customer data was used in training — when it needed examples of a task, it had its own lawyers and tax professionals produce them rather than mine client work product. Owning the weights also means owning deployment, hosting and audit trails, which is exactly the conversation general counsels have been having about sending privileged documents through third-party APIs. The company has engaged several outside security firms to validate the controls around serving the model.

Reaction

The legal technology press treated the launch as a milestone rather than a surprise — Thomson Reuters CEO Steve Hasker had talked openly about building an in-house model back in June, and preliminary benchmark figures surfaced earlier in August. What was new was the confirmation that the thing actually shipped, and the price tag attached to it.

Early outside testing has been limited but not entirely absent. Jonathan H. Choi, a professor at Washington University School of Law, ran Thomson against ChatGPT and Claude using hard questions from his Corporate Tax class. In a statement provided by the company, he said all three answered correctly but that he preferred Thomson's responses overall, singling out the links to treatises that made the answers easier to verify. That is a small sample and a company-supplied quote, and it should be read as such.

The internal benchmark, which Thomson Reuters calls Deep Research, is more interesting for what it measures than for the scores. Queries were drafted by practicing experts, who also wrote the criteria a good answer should meet. Responses were graded on completeness and on factuality — specifically, whether the citations offered actually supported the claims. Andrew Bean, who runs the evaluations team, framed it this way: a model can give a correct answer, but the better outcome is a correct answer with proof attached. Against GPT 5.4 and Claude Sonnet 5 with web access only, Thomson landed inside the pack but not ahead of it. With access to Thomson Reuters content, it moved past both.

That result is honest about its own limits. It says specialization plus proprietary tools beats general intelligence plus the open web on legal research questions — which is close to tautological, and also close to the only claim the company needs. Independent validation has not happened yet.

What's Next

A technical report is due imminently, covering standard public benchmarks, legal-specific evaluations and human testing. Thomson Reuters has also started handing the model to legal and AI academics for direct evaluation, and it is releasing a small open-weight version on Hugging Face under a non-commercial academic license. Hron described the open-weight release as a way for "anybody in the world to pick it up and critique and validate or invalidate any aspects of what we're saying" — an unusually exposed position for a company whose benchmark claims are the story.

A developer portal is in progress, with API keys, configurable parameters and sample documentation. Hron demonstrated an early build during the media briefing and was careful to call it very early. Separately, the company has begun conversations with large law firms and corporations about licensing Thomson directly, including firms that want to fine-tune it on their own matter files. "We built Thomson as infrastructure for Thomson Reuters," Hron said. "What we are beginning to see is that it could also become infrastructure for others."

The largest number in the announcement is the one about what has not been used yet: Thomson has been trained on less than 10 percent of the company's available content. The company is explicit that the next step is not simply pouring in the other 90 percent, but figuring out which content and which product activity convert into useful training signal. Extending Thomson models across the broader legal and tax portfolio is the stated destination.

Closing Thoughts

The obvious question is whether a media and information company can keep pace with laboratories that iterate every few months. Hron's answer is that it does not have to — improvements in open models become his next foundation, so the frontier labs' progress is an input rather than a threat. "Thomson does not need to keep pace with the frontier of general intelligence across all dimensions," he said. "Thomson needs to set the frontier of intelligence for legal." Whether that holds depends on whether the gap between a specialist and a very good generalist stays wide enough to be worth $40 million.

What is harder to argue with is the shape of the bet. Thomson Reuters spent a century and a half building an editorial asset that nobody can replicate, and it has now found a way to convert that asset into something that behaves like software. The model is the delivery mechanism; the 150 years of curation is the actual product. Hron put it plainly: he does not see owning a model that embodies the company's expertise as non-core work, because AI is simply a new mechanism for delivering expertise.

If Thomson holds up under outside scrutiny, the interesting consequence is not for law firms. It is for every other business sitting on a deep, narrow, painstakingly maintained corpus — medical publishers, engineering standards bodies, financial data providers — who have spent two years assuming the only move available was retrieval on top of somebody else's model. A $450,000 training run reframes that assumption considerably.

한글 요약

톰슨로이터가 8월 24일 자체 대규모 언어 모델 '톰슨(Thomson)'을 공식 출시했습니다. 처음부터 만든 파운데이션 모델은 아니고, 오픈 웨이트 모델(최근 버전은 Qwen 3.5)을 기반으로 웨스트로·프랙티컬로·체크포인트·로이터의 자체 콘텐츠를 학습시켜 법률 전문 모델로 만든 것입니다. 2년간 인력과 컴퓨팅에 약 4천만 달러를 썼지만, 이번에 출시된 버전의 최종 학습 실행 비용은 약 45만 달러에 그쳤다고 밝혔습니다. 프런티어 랩들이 수십억 달러를 언급하는 것과 비교하면 자릿수가 다른 숫자입니다.

톰슨은 법률 AI 어시스턴트 CoCounsel Legal의 표 형식 문서 검토(Tabular Analysis) 기능에 먼저 투입됩니다. CoCounsel 자체는 여전히 멀티 모델 구조를 유지하며, 톰슨이 확실히 유리한 작업에만 기본값으로 쓰입니다. 자체 벤치마크에서는 웹만 사용할 경우 GPT 5.4·Claude Sonnet 5와 비슷한 수준이었지만, 톰슨로이터 콘텐츠에 접근하면 두 모델을 앞섰습니다. 다만 아직 외부 독립 검증은 이뤄지지 않았고, 회사는 이번 주 기술 보고서 공개와 함께 소형 버전을 허깅페이스에 비영리 학술 라이선스로 공개할 예정입니다.

이 소식의 핵심은 법률 시장보다 그 바깥에 있습니다. 150년간 축적한 편집 자산이 있다면, 이제 남의 모델을 빌려 쓰는 대신 자기 모델을 직접 소유하는 선택지가 현실적인 가격대로 내려왔다는 뜻이기 때문입니다. 고객 데이터를 학습에 쓰지 않았다는 점, 배포와 거버넌스를 직접 통제한다는 점도 기업 고객에게는 비용만큼 중요한 조건입니다. 톰슨로이터는 아직 보유 콘텐츠의 10% 미만만 학습에 사용했다고 밝혔습니다.

참고 / 출처: LawSites, SiliconANGLE