Thomson is Thomson Reuters’ first proprietary large language model, built specifically for legal, tax and regulatory work rather than general-purpose chat. Announced on August 24, 2026, it cost about $40 million in talent and compute to develop, uses Thomson Reuters’ proprietary professional content and is being deployed first inside CoCounsel Legal’s Tabular Analysis.
That makes the launch more important than another model benchmark. Thomson Reuters is testing a different AI strategy: own the model, specialize it deeply, keep customer data out of training, and use outside frontier models only where they remain the better tool.
What is Thomson?
Thomson is a proprietary LLM developed by Thomson Reuters for professional workflows. The company says it began with a strong open-source foundation and then applied mid-training and post-training using its legal, tax, regulatory and news expertise, including content from Westlaw, Practical Law, Checkpoint and Reuters.
Thomson Reuters says hundreds of subject-matter experts helped shape training objectives and evaluations. That matters because the model is not being positioned as a replacement for ChatGPT, Claude or Gemini for everyday questions. It is designed for work where citations, domain knowledge and auditability carry more weight than broad conversational ability.
Why did Thomson Reuters build its own model?
The practical reason is control. Building a proprietary model lets Thomson Reuters decide how the model is trained, where it runs, how it behaves and which professional workflows it is optimized for. It can also reduce dependence on the pricing and product roadmaps of outside AI labs.
The company says the total investment in Thomson was $40 million across talent and compute. That is far below the multibillion-dollar infrastructure spending associated with the largest frontier-model programs, although the comparison is not apples-to-apples: Thomson starts from an open-source foundation and specializes it for narrower professional domains.
This follows a broader enterprise pattern discussed in BrandLigo’s AI agents for business guide: organizations increasingly get value by narrowing AI to a well-defined workflow, controlled data and measurable outputs rather than trying to automate everything with one model.
Where will lawyers actually use Thomson?
The first customer-facing deployment is Tabular Analysis in CoCounsel Legal. Thomson Reuters says the feature can review up to 10,000 documents against as many as 100 questions and return results in a filterable table with answers traceable to source material.
Thomson will not replace every model inside CoCounsel. Thomson Reuters explicitly describes CoCounsel Legal as multi-model: Thomson is used where the domain-specific model has an advantage, while other leading models can continue to power other tasks.
Thomson vs. ChatGPT, Claude and Gemini
| Area | Thomson | General frontier models |
|---|---|---|
| Primary audience | Legal, tax and regulatory professionals | Broad consumer and enterprise use |
| Training strategy | Open-source foundation plus proprietary professional specialization | Large general-purpose training programs |
| Standalone access | Not currently sold as a standalone model | Usually available through apps and APIs |
| First deployment | CoCounsel Legal Tabular Analysis | Chat, coding, research, agents and APIs |
| Customer data used for training | Thomson Reuters says no | Varies by provider and plan |
Thomson Reuters reports that Thomson performs competitively with leading frontier models on its evaluation suite and leads the models it tested on the difficult PrBench Legal Hard benchmark. Those are company-reported results, not a universal ranking of AI systems. Independent academic evaluation is underway, so buyers should treat the benchmark claims as promising evidence rather than a final verdict.
What does the $40 million figure really tell us?
The most useful lesson is not that frontier AI suddenly costs only $40 million. Thomson Reuters did not build a general-purpose frontier lab from scratch. It started from an existing open-source foundation, narrowed the problem and applied valuable proprietary data and expert feedback.
That suggests a potentially important path for other information-rich businesses. A company may not need to train the world’s largest model if its competitive advantage comes from specialized data, workflow integration, expert evaluation and a smaller model that is cheaper to operate.
Why proprietary data may matter more than model size
Thomson Reuters says less than 10% of its proprietary content has been used in Thomson’s training so far. The company argues that the biggest performance gains come from combining authoritative content with professional expertise during training, not merely retrieving documents after a user asks a question.
That is a meaningful distinction. Retrieval-augmented generation can give a general model access to a legal database at query time. Domain training tries to change how the model itself follows instructions and navigates specialized material. The two approaches can also be combined.
What about privacy and AI sovereignty?
Thomson Reuters says customer data is not used to train Thomson. It also frames ownership of the model as an AI-sovereignty advantage because it controls training, deployment and behavior rather than leaving those decisions entirely to a third-party model provider.
For US law firms and corporate legal departments, that does not eliminate due diligence. Buyers still need to understand retention, access controls, matter confidentiality, audit logs, model updates and the contractual terms of the product they actually use.
Can you buy Thomson directly?
No. Thomson Reuters says Thomson is not currently priced, sold or offered as a standalone model. Customers will encounter it first as a model layer inside CoCounsel Legal, beginning with Tabular Analysis.
That means a law firm evaluating the launch should compare the workflow and resulting work product rather than trying to compare a hypothetical Thomson API price with ChatGPT or Claude subscriptions.
What this means for the professional AI market
Thomson is evidence that the AI market may split into two complementary layers. General-purpose frontier models will continue to handle broad reasoning, coding and agentic work. Domain owners with valuable proprietary data may increasingly build or fine-tune smaller models for narrow, high-stakes tasks.
The same tension already appears in on-device AI, where smaller specialized models trade breadth for privacy and control. BrandLigo’s Meta Muse Glimmer guide looks at that trade-off from the local-computing side, while our Muse Code analysis shows how model economics are reshaping software tools.
Should US law firms change their AI strategy because of Thomson?
Not overnight. The sensible response is to evaluate Thomson inside the workflow where it is actually being deployed. Measure citation quality, review time, error rates, document throughput and how often lawyers need to correct the output.
The bigger strategic question is whether a domain-specific model produces enough improvement in trusted professional work to justify choosing it over a general model with access to the same documents. Thomson Reuters believes the answer is yes. Independent testing and real customer use will determine how broadly that holds.
Frequently asked questions
What is the Thomson AI model?
Thomson is Thomson Reuters’ proprietary large language model for legal, tax and regulatory work. It is built on an open-source foundation and specialized using Thomson Reuters content and professional expertise.
How much did Thomson cost to build?
Thomson Reuters says it invested $40 million in talent and compute to train Thomson. That figure covers its specialized model-development approach and should not be interpreted as the cost of building a general-purpose frontier AI lab from scratch.
Is Thomson available to the public?
Not as a standalone commercial model. Thomson Reuters says customers will first use it through Tabular Analysis in CoCounsel Legal. A small open-weight version is being made available for academic and non-commercial evaluation.
Does Thomson Reuters train Thomson on customer data?
Thomson Reuters says it does not use customer data to train Thomson.
Is Thomson better than ChatGPT or Claude?
Thomson Reuters reports competitive results against leading frontier models on professional and legal benchmarks, but those results do not mean Thomson is better for every task. It is deliberately specialized, and independent academic evaluation is still underway.
Will Thomson replace other models in CoCounsel?
No. CoCounsel Legal remains a multi-model product. Thomson Reuters says it will use Thomson where the specialized model provides an advantage and other leading models where they are better suited.
Sources: Thomson Reuters launch announcement; Thomson model overview; How Thomson was built; CoCounsel Legal announcement; and Reuters on the competitive legal-AI market.
About this article. Written and fact-checked by the BrandLigo editorial desk. Product, deployment and benchmark claims were checked against Thomson Reuters’ primary materials on August 25, 2026; company benchmark claims are labeled as such.