Thomson 1.0: Thomson Reuters' Own Legal-Tax LLM

The turning point of an information giant

For decades, Thomson Reuters sold the most valuable input for lawyers and accountants: organized, reliable, and up-to-date information. Westlaw, Practical Law, Checkpoint — foundations that structure legal and tax practice across multiple countries. On August 24, 2026, the company took a qualitatively different step: it announced Thomson 1.0, its first internally developed large language model (LLM), designed specifically for the legal and tax domains.

This is no longer a partnership with OpenAI. No longer a wrapper over GPT or Claude. It is a proprietary model, trained on the company's own content assets, built around a clear thesis: no one has legal and tax data as deep as Thomson Reuters — and that asset, converted into model intelligence, can be a generational competitive advantage.

US$ 40 million and a final training run costing US$ 450,000

The total investment declared by Thomson Reuters to develop Thomson 1.0 was US$ 40 million over approximately two years, covering talent and computing infrastructure. That figure includes the acquisition of Safe Sign Technologies, completed in 2024, which brought in part of the technical team responsible for the project.

The most revealing figure, however, came from CTO Joel Hron during the press briefing: the cost of the model's final training run was US$ 450,000 — a number self-reported by the company. For context, Thomson Reuters itself describes the US$ 40 million total as "a fraction" of the billions invested by frontier labs such as OpenAI and Anthropic. This is not modesty: it is a methodological bet. The company believes that extremely high-quality, sector-specific data can compensate for orders of magnitude in raw scale.

The model was led technically by Joel Hron, CTO, and Jonathan Schwarz, Head of AI Research. Hundreds of subject-matter experts participated in defining training objectives and evaluating results — a deliberate choice to blend engineering with domain expertise.

What is under the hood

Thomson 1.0 was built on an open-source or open-weight foundation. Thomson Reuters' official materials do not identify which base model was used — and two specialized press sources diverge on this point, citing, respectively, Qwen 3.5 (Alibaba) and Snowdon, developed at Imperial College London. Since the available reporting documents this contradiction without independent resolution, and the company's official materials do not specify, the point remains open.

What is documented is the fine-tuning fuel: Westlaw, Practical Law, Checkpoint, and Reuters — the company's proprietary content assets. As of launch, less than 10% of that total corpus had been used in training, according to Thomson Reuters' own disclosure. This means there is a substantial reserve of data not yet incorporated — which the company presents as a key driver of evolution for upcoming versions in the Thomson model family.

One critical point on privacy: customer data was not used in training without explicit consent, according to the company.

"Fiduciary-Grade AI": a proprietary standard — and a promise

Thomson Reuters coined the term "Fiduciary-Grade AI™" to describe its development standard, making a direct reference to the legal concept of duty of care. The choice is not accidental: the target audience — lawyers and accountants — operates under strict professional liability regimes. A model that hallucinates a piece of case law or misinterprets a tax provision is not merely inconvenient; it can generate legal liability.

The company states that Thomson 1.0 was evaluated against more than 100 legal-domain benchmarks and thousands of specific use cases. On the PrBench Legal Hard benchmark, the model recorded a score of 0.352, which Thomson Reuters describes as superior to all frontier models tested.

As for comparisons with competing models, available sources diverge in their claims: one asserts that Thomson performs competitively with Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro; another claims competitive performance with Claude Opus 4.8 and superior performance to GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro across a series of legal and professional benchmarks. Both versions are self-reported by Thomson Reuters and there is no consensus between sources. A mandatory caveat: all of these figures are self-reported by the company. Independent academic benchmarking was underway with partner institutions at the time of launch, but results had not yet been published. Any critical analysis must account for this limitation before treating the comparisons as definitive.

What is available — and what is not yet

Here lies the most important detail for anyone evaluating immediate adoption: Thomson 1.0 is not available as a standalone product as of the launch date. The company's official FAQ is direct: "No. We do not currently price, sell, or offer Thomson as a standalone model."

The first commercial deployment will be the Tabular Analysis feature within CoCounsel Legal, in a "forthcoming version" of the product — not yet live as of August 24, 2026. This feature enables the review of up to 10,000 documents cross-referenced against up to 100 questions simultaneously, with each answer traceable to its source. For legal departments handling due diligence, large-scale contract review, or tax audits, the potential for time compression is evident.

A smaller open-weight version is available on Hugging Face under a non-commercial academic license — accessible to researchers and technical teams who want to experiment with the model. A developer API portal exists, but CTO Hron described it as "still very early" at the time of the announcement. Preliminary conversations with large law firms and corporations about direct licensing and fine-tuning with proprietary data are underway.

One relevant technical limitation: Thomson 1.0 does not query real-time updated databases by default — unlike CoCounsel, which has native access to Thomson Reuters' proprietary content. The company is evaluating the addition of real-time retrieval in future versions.

What this changes in practice for businesses

For law firms, corporate legal departments, and tax consulting practices, Thomson 1.0 signals a trend that goes well beyond Thomson Reuters: sectoral information providers are no longer passive clients of AI — they are becoming owners of verticalized models.

This has direct implications. Tools like Tabular Analysis — when available — can compress into hours due diligence tasks that today take weeks of junior associate work. More than speed, the traceability of every answer back to its source is what makes this type of tool defensible within a regulatory environment.

For technology and operations managers at legal and accounting firms, the question is no longer "should we use AI?" but rather "which AI layer, with what data, and with what level of auditability?" Thomson 1.0 is the answer from a company betting that the right answer starts with data — not with model size.

The Thomson model family is just getting started. With less than 10% of the proprietary corpus used so far, what comes next depends less on computing power and more on curation. That is the game Thomson Reuters is playing.

Sources

• Thomson Reuters investe US$ 40 milhões e lança modelo ...
• Thomson Reuters Launches Thomson, Its Own Proprietary ...
• Thomson Reuters Leverages its World-Class Data Assets ...
• Thomson Reuters lança modelo de IA próprio por US$ 40 milhões Por Investing.com
• Thomson Reuters launches $40M legal LLM inside CoCounsel
• Thomson Reuters launches proprietary legal LLM "Thomson"