Thomson Reuters Builds Its Own AI Model for $40 Million

Thomson Reuters Is Taking a Different Route to Frontier AI

Thomson Reuters has launched Thomson, its first proprietary large language model, in a move that highlights a changing strategy in enterprise artificial intelligence. Instead of trying to compete with the biggest general-purpose models through enormous training budgets, the company started with an open-source foundation and focused its investment on specialized professional knowledge, training, and domain expertise. 

The company says it invested about $40 million in talent and computing to develop Thomson. That is a very different scale from the multibillion-dollar infrastructure investments associated with the largest frontier AI labs. Thomson Reuters argues that the approach can produce a highly capable model while giving the company direct control over the technology and its operating costs.

What Makes Thomson Different?

The important part of the announcement is not simply that Thomson Reuters now owns an LLM. The more interesting decision is where the company concentrated its effort.

Thomson begins with a strong open-source foundation and is then trained and refined using decades of Thomson Reuters expertise and proprietary material from areas including Westlaw, Practical Law, Checkpoint, and Reuters. Hundreds of subject-matter experts were involved in designing training objectives and evaluating the model. 

Thomson Reuters says less than 10% of its available proprietary content has been used in training so far. The company views its remaining specialized content and expertise as a potential source of further improvement rather than simply treating more data as the answer. 

Why Specialized AI May Matter More Than Model Size

For general-purpose chatbots, model scale can be an important part of the capability race. Professional software has a different requirement: the system needs to understand specialized terminology, follow complex instructions, work with trusted information, and produce results that professionals can inspect and rely on.

That changes the economics of the problem. A company that already owns valuable domain data and understands a particular workflow may not need to build the largest model in the world. It may get more value by taking a capable foundation model and specializing it around the work its customers actually perform.

Thomson Reuters CTO Joel Hron described this as a shift away from treating bigger models and more compute as the only route to better AI. The company's broader argument is that enterprises can gain an advantage by knowing when to use general-purpose intelligence and when specialized intelligence is more appropriate. 

What Thomson Reuters Says About Performance

Thomson Reuters says its early evaluations put Thomson on par with leading frontier models across a range of tasks. Its model information page reports competitive results against models including Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, while also reporting a leading result on the PrBench Legal Hard benchmark.

Those results are promising, but they should be read carefully. The headline performance numbers come from evaluations presented by Thomson Reuters, while the company says independent academic benchmarking is still underway. That distinction matters because benchmark results can depend heavily on the dataset, task design, model configuration, and evaluation methodology. 

External coverage has also highlighted this limitation. One independent analysis argued that the $40 million figure needs context and that publicly available evidence is not yet the same as broad independent proof that Thomson consistently beats the leading general-purpose models. 

The First Real-World Test Is CoCounsel Legal

Rather than launching Thomson as another general chatbot, Thomson Reuters is putting it to work where its specialization should matter most. The first deployment is inside Tabular Analysis in CoCounsel Legal, where the model is intended to help with high-volume, structured document-review work. 

This is a useful test because real-world professional AI is ultimately judged by more than benchmark scores. Lawyers and corporate legal teams care about accuracy, citations, consistency, explainability, privacy, and how reliably a system fits into an existing workflow.

Thomson Reuters is also keeping CoCounsel Legal multi-model. Instead of forcing every task through Thomson, the company says it will use Thomson where its specialized capabilities provide the clearest advantage and continue using other leading models elsewhere. 

Why AI Sovereignty Is Becoming a Bigger Issue

Owning the model gives Thomson Reuters more control over how its AI is trained, operated, and integrated with its proprietary tools. This can matter for customers handling sensitive legal, tax, regulatory, and corporate information.

The company describes this broader concept as AI sovereignty: organizations increasingly want to know where their AI runs, how it was trained, what behaviors and biases may exist inside it, and how customer information is handled.

That does not mean an internally developed model automatically becomes more trustworthy. Ownership provides additional control, but trust still depends on testing, evaluation, security, governance, data practices, and the quality of the systems built around the model.

The Bigger Lesson for Enterprise AI

Thomson Reuters' launch points toward a broader change in enterprise AI strategy. Companies may increasingly stop asking only, "Which model is the smartest?" and start asking, "Which parts of our AI stack should we control ourselves?"

For some businesses, the answer will remain a hosted general-purpose model. Building and operating a proprietary foundation model is not automatically cheaper or better. For organizations with valuable proprietary data, specialized workflows, strict privacy requirements, and enough resources to invest in model development, however, a smaller purpose-built model could make strategic sense.

The most important advantage may therefore be the combination of model + proprietary data + domain expertise + workflow integration, rather than model size alone.

What This Means for Other Software Companies

Thomson Reuters is not the only company that can take this approach. Other software businesses with large proprietary datasets and deep industry knowledge could consider adapting open models rather than building every component from scratch. AI Business described the launch as a potential example for other SaaS companies looking to turn their specialized data into an AI advantage. 

The economics will vary widely, though. Thomson Reuters already had decades of professional content, established software products, subject-matter experts, and customer workflows. A company without those assets would have a much weaker reason to make the same investment.

What We Still Need to See

The next stage of Thomson's development will be more revealing than the launch announcement itself. Independent evaluations, customer results, error rates, citation quality, operating costs, and performance across different professional tasks will show whether the model's specialization produces a durable advantage.

Thomson Reuters has said it plans to make a small version available as an open-weight model for academic and non-commercial use, while continuing external evaluation with academics. That could provide additional evidence beyond the company's own internal testing. 

The company also plans to expand Thomson across its legal and tax portfolio and develop additional sovereign AI options. 

Frequently Asked Questions

What is Thomson?

Thomson is Thomson Reuters' proprietary large language model, designed primarily for professional work such as legal, tax, and regulatory applications.

How much did Thomson Reuters spend on the model?

Thomson Reuters says it invested approximately $40 million in talent and computing to develop Thomson. 

Is Thomson a general-purpose chatbot?

No. Thomson is positioned as a purpose-built model for professional applications rather than a direct replacement for general consumer chatbots.

Where is Thomson being used first?

The first deployment is planned for Tabular Analysis within CoCounsel Legal, targeting structured document-review work for law firms and corporate legal departments. 

Does Thomson Reuters claim Thomson beats leading AI models?

The company says Thomson is competitive with leading frontier models on several evaluations and reports strong legal-benchmark performance. However, broader independent academic evaluation is still underway, so those claims should be treated as company-reported results until independently replicated. 

The Real AI Race May Be About What Companies Own

Thomson Reuters' new model does not prove that specialized AI will replace frontier models. It does show why the enterprise AI race is becoming more complicated than simply building the biggest model.

General-purpose models will continue to push the boundaries of broad intelligence. But companies with valuable proprietary information may increasingly want an intelligence layer that understands their data, follows their standards, fits their workflows, and remains under their control.

That makes Thomson an interesting development not because a $40 million model has suddenly changed the entire AI industry, but because it offers a practical alternative to the assumption that every company must rent intelligence from the largest AI labs. The next competitive advantage may belong to organizations that know which intelligence to build, which intelligence to buy, and how to combine the two.