Reuters invets $40 Million To Build Its Own AI Model Instead Of Renting From OpenAI

Thomson Reuters is betting that owning a specialized artificial intelligence model can create more long-term value than continuously paying frontier AI companies for access to their models. The legal, tax and professional-information giant has invested about $40 million in developing its proprietary large language model, Thomson, over more than two years, seeking greater control over AI costs, performance, intellectual property and the technology used to power its professional products.

The strategy does not mean Thomson Reuters is abandoning OpenAI or Anthropic. The company has continued to work with Anthropic, including an expanded partnership for CoCounsel Legal announced in May. Instead, Thomson Reuters is building its own model for tasks where its decades of proprietary legal, tax, accounting and news content can provide a competitive advantage.

Thomson Reuters Invests $40 Million In Its Own AI Model

Thomson Reuters said it invested $40 million in training Thomson, including talent and computing resources. The company emphasized that the figure represents the broader development effort rather than simply the cost of one training run.

The project began as an internal effort more than two years ago and has evolved into a proprietary AI system designed specifically for professional work. Thomson Reuters says the model is fully owned and controlled by the company.

Thomson AI Investment At A Glance

Metric	Details
Company	Thomson Reuters
Proprietary model	Thomson
Development investment	About $40 million
Development period	More than two years
Primary focus	Professional AI
Key sectors	Legal, tax, accounting, regulatory
Starting foundation	Open-source model
Ownership	Fully owned and controlled by Thomson Reuters
Initial production use	CoCounsel skills
Major external AI partner	Anthropic

The investment is relatively small compared with the multibillion-dollar infrastructure spending associated with leading frontier AI laboratories. Thomson Reuters says its approach was to start with an existing open-source foundation and specialize the model using proprietary data and expert knowledge.

Why Thomson Reuters Wants To Own Its AI

The company’s reasoning is straightforward: Thomson Reuters already owns valuable professional information that general-purpose AI companies do not have in the same form.

The company has accumulated approximately 175 years of authoritative data across legal, tax, accounting and news businesses. Its major information assets include Westlaw, Practical Law, Checkpoint and Reuters. It also employs thousands of subject-matter experts who help determine what constitutes accurate professional work.

That creates an opportunity to build a model around knowledge and workflows that are highly specialized rather than trying to compete with general-purpose models on every possible task.

Thomson Reuters’ AI Assets

Asset	Strategic Value
Westlaw	Legal research and case information
Practical Law	Legal guidance and workflows
Checkpoint	Tax and accounting content
Reuters	Global news and information
Subject-matter experts	Human evaluation and professional judgment
Professional software	Real-world AI deployment environments
Customer workflows	Domain-specific training and testing

The company’s argument is that these assets can make a smaller, specialized model highly effective for professional applications.

From Renting AI To Owning AI

Thomson Reuters CTO Joel Hron has compared the strategy to buying rather than renting a home.

Using an external model means paying an AI provider for access to its technology and remaining exposed to that provider’s pricing, roadmap and infrastructure decisions. Building an internal model requires substantial upfront investment but gives the company greater control over the underlying technology.

EXTERNAL AI MODEL

OpenAI / Anthropic
        ↓
API Access
        ↓
Usage Fees
        ↓
Thomson Reuters
        ↓
Customer Product


PROPRIETARY AI MODEL

Thomson Reuters Data
        +
Expert Knowledge
        ↓
Thomson Model
        ↓
Company-Owned IP
        ↓
Customer Products

The model does not eliminate external AI costs entirely, but it potentially reduces dependence on third-party frontier models for selected workloads.

Thomson Is Built For Professional Work

Thomson Reuters is not positioning Thomson as simply another general-purpose chatbot.

The company designed it around professional tasks where accuracy, citations, consistency and auditability matter. Initial production deployment is focused on high-volume structured document review within CoCounsel, with further integration across the legal and tax portfolio planned.

This distinction is important because professional AI can have different requirements from consumer chatbots.

A lawyer reviewing a contract, for example, may value traceable sources and reliable document analysis more than the ability to generate entertaining or creative responses.

Where Thomson Is Initially Targeted

Application	Relevance
Document review	High-volume structured analysis
Legal research	Specialized legal knowledge
Tax analysis	Professional tax content
Accounting	Domain-specific information
Regulatory work	Accuracy and auditability
CoCounsel	AI-assisted professional workflows

Thomson Reuters says the model has been trained using authoritative proprietary content and evaluated with substantial involvement from subject-matter experts.

The Model Started With An Open-Source Foundation

Thomson Reuters did not build Thomson entirely from scratch.

The company began with a strong open-source model foundation and then applied its own training, mid-training and post-training techniques. According to reporting on the project, the underlying work involved a version of Alibaba’s Qwen model that was adapted before Thomson Reuters applied its proprietary data and professional training process.

The company says the resulting Thomson model is what matters commercially because it has been substantially specialized for its own professional applications.

Thomson Development Process

Open-Source Foundation
          ↓
Safety / Alignment Work
          ↓
Proprietary Thomson Reuters Data
          ↓
Mid-Training
          ↓
Expert-Guided Post-Training
          ↓
Agentic Reinforcement Learning
          ↓
Professional Evaluations
          ↓
THOMSON

Thomson Reuters worked with organizations including Imperial College London, DatologyAI, Lambda and Together AI during development.

Proprietary Data Is The Core Advantage

The company’s strategy highlights an increasingly important question in enterprise AI: Does a company need the world’s biggest model, or does it need the model that understands its business best?

For Thomson Reuters, proprietary information may be more valuable than raw model size.

A general-purpose AI model can know about law, tax and accounting, but Thomson Reuters has decades of curated information and professionals who work with those subjects every day.

That provides a potential competitive advantage that cannot easily be replicated simply by increasing computing power.

General AI Vs Specialized AI

Factor	General Frontier Model	Thomson
Primary objective	Broad intelligence	Professional work
Training advantage	Massive general datasets	Proprietary professional content
Domain expertise	Broad	Deep
Legal workflows	General capability	Purpose-built
Citations	Model-dependent	Key design priority
Ownership	External provider	Thomson Reuters
Cost structure	Usage/provider dependent	Company-controlled
Customization	API/model options	Internal development

The strategy is particularly relevant in industries where incorrect information can create significant financial or legal consequences.

Thomson Reuters Says Thomson Can Compete With Frontier Models

Thomson Reuters released benchmark results in July showing Thomson performing competitively with leading frontier models.

The company said its model performed competitively with Claude Opus 4.8 and ahead of GPT-5.5, Claude Sonnet 5 and Gemini 3.1 Pro across a range of legal and general benchmarks.

Thomson Reuters also highlighted external academic evaluations. One law professor said Thomson produced responses he preferred to those from ChatGPT and Claude on challenging corporate tax questions, particularly because of its links to legal treatises.

These are company-reported results and should therefore be viewed in that context rather than as an independent industry ranking.

The Cost Advantage Could Be Significant

One of the strongest arguments for Thomson’s strategy is economics.

Frontier AI companies spend enormous amounts on computing infrastructure and model training. Thomson Reuters says Thomson was developed at a fraction of the cost of comparable frontier models.

The company also claims the model can be operated at a fraction of the inference cost of typical frontier models.

AI Economics Comparison

Cost Factor	Frontier AI Approach	Thomson Reuters Approach
Initial investment	Extremely high	About $40 million reported
Model foundation	Built at frontier scale	Open-source foundation
Training data	Broad/general	Proprietary professional data
Inference	External or large-scale infrastructure	Designed for efficiency
Model ownership	AI provider	Thomson Reuters
Pricing exposure	Third-party pricing	Greater internal control
Customization	Provider-dependent	Company-controlled

The economics become more attractive when an AI model handles millions of repetitive professional tasks. Lower inference costs can potentially improve margins as usage scales.

Thomson Reuters Is Not Cutting Ties With Anthropic

The company’s decision to build Thomson should not be interpreted as an outright rejection of external AI providers.

In May, Thomson Reuters and Anthropic expanded their partnership to connect Claude with CoCounsel Legal. The agreement brought CoCounsel Legal into Claude workflows through an MCP integration.

Thomson Reuters has also previously evaluated models from OpenAI, Anthropic and Google as part of its multi-model strategy.

This suggests the company is pursuing a hybrid AI strategy rather than betting everything on one model.

                THOMSON REUTERS AI
                       │
          ┌────────────┴────────────┐
          ↓                         ↓
     THOMSON MODEL            EXTERNAL MODELS
          │                         │
          ↓                         ↓
 Specialized Tasks          General / Partner Tasks
          │                         │
          └────────────┬────────────┘
                       ↓
                 CoCounsel +
             Professional Products

The proprietary model can handle workloads where Thomson Reuters’ data provides an advantage, while external models can remain useful for broader capabilities.

AI Ownership Could Reduce Vendor Dependence

Enterprise customers are increasingly concerned about dependency on a small number of AI providers.

If an AI company changes pricing, model access, usage limits or product priorities, customers built heavily around that provider may have limited alternatives.

Owning an internal model gives Thomson Reuters another option.

The company can control its model roadmap and optimize it for the specific tasks its customers require rather than waiting for a general-purpose AI provider to prioritize those capabilities.

Potential Strategic Benefits

Benefit	Potential Outcome
Model ownership	Greater technology control
Lower inference costs	Potential margin improvement
Proprietary training	Stronger domain specialization
Internal roadmap	Faster product customization
Reduced vendor dependence	Greater negotiating flexibility
Data integration	Better use of proprietary information
Expert feedback	More targeted model improvement

The approach could become increasingly attractive for other large enterprises with substantial proprietary datasets.

Thomson Reuters Is Turning Into An AI Company

The Thomson project represents a broader transformation at the company.

Thomson Reuters has traditionally been a provider of professional information, software and workflow tools. AI is increasingly becoming embedded directly into those products.

The company reported that 32% of its contract value involved generative AI in the second quarter of 2026, up from 30% in the previous quarter, according to Reuters.

That makes the economics of AI increasingly important to Thomson Reuters’ overall business.

Thomson Reuters AI Strategy

Authoritative Data
        +
Expert Knowledge
        +
Professional Software
        +
AI Models
        ↓
AI-Powered Professional Workflows
        ↓
Higher Customer Value
        ↓
Recurring Software Revenue

Owning the model could allow the company to capture more of that value rather than simply passing a portion of AI-related revenue to external model providers.

A Small Model Can Be Valuable If It Knows The Right Things

The Thomson project also challenges the assumption that the biggest AI model will always be the best commercial model.

For professional applications, a smaller system trained and evaluated specifically for a narrow set of high-value tasks can potentially outperform larger general-purpose models on those tasks while costing less to operate.

This is particularly relevant in legal and tax environments where customers often need accuracy, references and consistency rather than unrestricted general intelligence.

Thomson Reuters’ approach can therefore be summarized as:

specialization over scale.

The Bigger Picture

Thomson Reuters’ $40 million investment reflects a broader shift in enterprise AI economics. Companies that own valuable proprietary data are increasingly asking whether they should simply rent intelligence from frontier AI providers or build specialized models that convert their own data and expertise into proprietary technology.

Thomson Reuters’ answer is to do both. The company continues working with Anthropic and evaluating external models while building Thomson for workloads where its proprietary content, professional expertise and customer workflows can provide an advantage.

The strategy is particularly notable because the company did not spend billions attempting to reproduce a frontier model from scratch. Instead, it started from an open-source foundation and invested in specialization, training, expert evaluation and infrastructure. Thomson Reuters says the resulting model is fully owned and controlled by the company and operates at a fraction of the cost of comparable frontier systems.

Looking Ahead

The biggest test for Thomson will be whether its proprietary model can move beyond benchmarks and deliver measurable improvements in real customer workflows. The company has begun deploying Thomson for high-volume structured document review in CoCounsel and plans further integration across its legal and tax portfolio. If the model can reduce inference costs while maintaining high accuracy and citation quality, the economics of owning specialized AI could become increasingly compelling.

For the wider AI industry, Thomson Reuters offers a potentially important enterprise model: companies do not necessarily need to compete with OpenAI or Anthropic at the absolute frontier to own valuable AI technology. Businesses with deep proprietary datasets, expert knowledge and specialized workflows may be able to build smaller, cheaper systems that are highly competitive where it matters most. That could lead to a more fragmented AI market in which frontier models remain important, but thousands of domain-specific models capture an increasing share of enterprise workloads