Thomson Reuters Builds a Frontier Legal Model with DatologyAI Data Curation

“For years, the AI industry has treated scale as the answer: bigger models, more compute, more money. Thomson shows there is another path: start with a strong foundation, specialize it deeply for the work that matters, and you can build intelligence that is highly capable, far more efficient, and entirely under your control. We think that changes the economics of professional AI.”

Joel Hron, CTO, Thomson Reuters

Results

• Competitive with the closed frontier - Thomson-1 outperformed Gemini 3.1 Pro, GLM 5.2, GPT-5.4, and Sonnet 5 on overall benchmark performance, and is competitive with Claude Opus 4.8.
• Outperforms GPT-5.6 Sol on complex legal workflows - Thomson-1 was preferred in blind expert evaluations over frontier models on legal tasks, like GPT-5.6 Sol, by 57% to 33%.
• ~100x lower cost per task - Thomson-1 offered comparable performance to GPT-5.6 Sol on legal agentic tasks, at a fraction of the cost to run.
• A ~$450K final training run - Training Thomson-1 cost a small fraction of what a frontier-quality model is supposed to cost.

Thomson Reuters holds one of the largest proprietary legal and news archives in the world. The challenge was turning that archive into a model Thomson Reuters could fully own, control, and run efficiently, without sacrificing the performance professionals expect from frontier AI. DatologyAI partnered with the AI research team at Thomson Reuters to support mid-training data curation as part of the overall model development process. The models and overall results are the product of Thomson Reuters’ full development effort.

From Proof of Concept to Production

Thomson Reuters started the journey toward its own intelligence for professional AI with key advantages: proprietary legal and news data no competitor can replicate, plus an experienced ML team and a post-training pipeline already built on that data with domain experts. The pipeline was working well but was reaching its limit. At the same time, the cost to run complex professional tasks on frontier models was rising as work became more agentic and heavy on multi-step reasoning. To keep improving the model for real professional work and control costs, the Thomson Reuters AI team realized they needed a more specialized model trained using more of their proprietary data and domain expertise. Rather than building critical data curation and synthetic data capabilities from scratch, they partnered with DatologyAI. 

To start, Thomson Reuters and DatologyAI ran a focused proof of concept applying mid-training data curation to Thomson Reuters' proprietary legal data. That work produced a clear signal: curated mid-training data made post-training more than twice as effective on Thomson Reuters' private legal evaluations, at a token budget under 1% of the base model's pretraining budget. This also had a bigger implication, revealing that the limit on post-training gains was the quality of the data used during model training to teach the model new skills, and that better data quality raises the ceiling on what post-training can do.

Thomson Reuters took that finding and built on it. On August 24, 2026, the company launched Thomson-1, its first proprietary large language model, developed in-house with Imperial College London, DatologyAI, and Lambda as technical partners. 

Thomson-1 was built through a Continual Learning approach: starting from a strong open-weight foundation and layering domain-specific mid-training, preference optimization, and reinforcement learning, rather than training a new model from scratch. The team at DatologyAI worked closely with Thomson Reuters on the data curation behind the mid-training stage, the same phase validated in the earlier proof of concept, now applied at production scale.

“A really nice finding we made together with DatologyAI, by looking at results and planning experiments jointly, is that we unlocked mid-training on instruction-tuned models. It's an unusual, unorthodox thing to do. The typical pipeline does mid-training on a base model and applies all the post-training afterward, which means losing the post-training that's already been done. Given how many gains are in post-training nowadays, you really want to keep it.” — Jonathan Richard Schwarz, Head of AI Research, Thomson Reuters

Curating a Frontier-Scale Mid-Training Dataset

Using DatologyAI, the team curated a 200B-token mid-training dataset from a corpus of permissively licensed public data and Thomson Reuters' proprietary content: decades of news, contracts, case law, statutes, regulations, and practitioner guidance from Westlaw, Practical Law, Checkpoint, and Reuters. The combined public and proprietary source pool exceeded 19T tokens; the final dataset kept under 2% of it, selected and enhanced through DatologyAI's data curation pipeline of source ingestion, quality and taxonomy classification, synthetic data generation, and algorithmic mixing. The entire pipeline ran inside Thomson Reuters’ own cloud environment, and the data never left their account.

The mid-training mix split roughly evenly across three components: curated proprietary documents, synthetic rephrasings of that content generated with an adaptation of DatologyAI's BeyondWeb method, and curated general-capability replay data designed to protect and enhance skills the model had already learned. In ablation studies, swapping a publicly available replay dataset for DatologyAI's curated replay mixture improved coding performance by an average of 5.4 percentage points across three benchmarks and reading comprehension by up to 7.2 percentage points, showing that curation quality on replay data has real, capability-specific effects.

An Enterprise Model That Competes with the Frontier

The Thomson-1 model delivers frontier-quality results for professional work, but at a fraction of the cost per task of frontier models, which is a huge win for the legal teams using AI for complex tasks that have become increasingly expensive with rising complexity and costs for frontier model APIs.

Figure 1: Expert Preferences for Thomson-1.0-Large vs. External Models

Preferred by Experts

The most telling result came from testing with human experts. In a blind evaluation, subject-matter experts judged more than 3,000 real legal and general conversations without knowing which system produced each answer. Thomson-1 ran as a complete system, connected to Reuters news and Thomson Reuters’ legal databases, against five frontier systems with broad web access. On legal tasks, the experts preferred Thomson-1 over every frontier system tested, including by 57% to 33% over GPT-5.6 Sol. On the complex legal workflows that professionals actually run, Thomson-1 showed that it can beat the best frontier models when it matters most.

Figure 2: Average Scores Across Benchmarks

Better Performance

Thomson-1 beat all the frontier models on legal domain average and on instruction-following and general agentic tasks. On benchmarks across legal, tax, journalism, general, and safety, Thomson-1.0-Large finished second only to Claude Opus 4.8, and scored better than Gemini 3.1 Pro, GLM 5.2, GPT-5.4, and Sonnet 5. Although mid-training focused on legal and professional content, Thomson-1 retained its broad capabilities and gained ones it was never explicitly trained for, like instruction following, with almost none of the forgetting typically associated with narrow domain adaptation.

“One of the exciting things about taking an instruct model at the start of mid-training is that you can leverage all the knowledge already put into it. Forgetting is a big issue, so we use two strategies to hold onto those capabilities: model merging, averaging the weights between the starting checkpoint and the model after mid-training, and replay data that emulates what the model has already seen.” — Jonathan Richard Schwarz, Head of AI Research, Thomson Reuters

Figure 3: Cost per Task Performance Along the Pareto Frontier

100x More Cost-Effective to Run

Thomson-1.0-Large matches the frontier's performance band at a fraction of the cost, while the models that match its score cost about 100x more per task to run. That matters to Thomson Reuters' customers because professional work is high volume and increasingly agentic. When each task costs a fraction of the frontier price, high-volume features stay economical, and Thomson Reuters can extend frontier-quality legal AI to more customers without the bill scaling with usage.

“Because it's trained on our own content and the Reuters news file, it's much more efficiently constructed than a large language model scouring the web. Relative to accessing a leading frontier model, we're going to have a significant cost advantage, which we'll be able to pass on to our customers.” — Steve Hasker, CEO, Thomson Reuters

Looking Forward

The team at Thomson Reuters built its first generation of models with a lean setup, including a team of fewer than three dozen people, a compute cluster capped at 368 GPUs, and a final training run of about $450K. While those numbers are astounding, what matters more than any single model run is that the company has built a model factory. Now they can combine any open-weights model they choose with their proprietary data to create a competitive, governable model they own. As new use cases emerge and stronger open-weights models become available, the team at Thomson Reuters is well positioned to build their next model using the reusable pipeline instead of starting over, which is a key advantage in the fast-moving AI marketplace.

It is also clear that the team has a lot of room to grow. This project drew on less than 10% of Thomson Reuters’ available proprietary content. Going forward, the teams at Thomson Reuters and DatologyAI will continue to work together to leverage a larger pool of data, further curate it to reduce hallucinations in high-stakes outputs, and create models tailored to a broader set of legal use cases. Thomson Reuters released Thomson-1 inside CoCounsel Legal's Tabular Analysis feature and the team also released a small, open-weight version on Hugging Face for academic and non-commercial evaluation, and these are just the first of many places these models and their successors will create an impact across Thomson Reuters products.

What began as a proof of concept validating the limits of post-training has become the data foundation for an ongoing program where Thomson Reuters owns its intelligence outright and continues to push the limits of what is possible for professional AI.

Additional Resources on Thomson-1

Watch the research podcast on Thomson-1 with the Thomson Reuters and Datology AI research teams - https://youtu.be/qsEFbYPJraM

Read the Thomson Reuters Tech Report on this project - https://huggingface.co/spaces/tri-fair-lab/publications/blob/main/Thomson_1_0_Technical_Report.pdf

Read the Thomson Reuters Press Release about Thomson-1: https://www.thomsonreuters.com/en/press-releases/2026/august/thomson-reuters-leverages-its-world-class-data-assets-to-launch-its-own-frontier-model

Watch Steve Hasker, CEO of Thomson Reuters, on Artificial Lawyer: https://www.artificiallawyer.com/2026/08/24/tr-ceo-steve-hasker-on-thomson-1-0/