Thomson Reuters Bets $40 Million on Its Own Legal AI to Curb Dependence on Anthropic

Thomson Reuters just took a calculated step away from its heavy reliance on outside AI providers. The information services giant launched Thomson-1, its first proprietary large language model. The move signals a shift in how one of the legal industry's biggest players thinks about building and owning the technology that powers its flagship products.

Announced this week, the model draws from an open-source foundation developed by Alibaba. Thomson Reuters adapted it through a process its chief technology officer described as realignment. The result? A system trained on decades of the company's own authoritative legal, tax and news content. Early tests show it holding its own against some of the most advanced general-purpose models on the market.

But don't mistake this for a full break from Silicon Valley's AI leaders. CoCounsel, Thomson Reuters' marquee AI assistant for lawyers, still leans primarily on Anthropic's Claude. The new model slots in for specific high-volume tasks where domain knowledge delivers a measurable edge. Joel Hron, the company's CTO, made the strategy plain.

Business Insider reported Hron's analogy. "Renting a house, you still have a roof over your head, and somebody's taking care of it, and it's great. But you're not building any equity that compounds into something valuable for you long term." The company spent roughly $40 million on compute, talent and specialized training to create that equity.

The numbers tell part of the story. Thomson Reuters used less than 10% of its vast proprietary corpus to train the model. Hundreds of subject-matter experts reviewed outputs. They identified failures. They refined the system to prioritize accuracy over pleasing responses. The approach stands in contrast to the race for ever-larger general models.

Thomson-1, also referred to as Thomson in company materials, builds on Snowdon. That variant stems from Alibaba's Qwen model. A joint team with Imperial College London spent months adapting it. They focused on ethical safeguards, de-biasing and safety. "There's nothing that necessarily ties us to Qwen," Hron told reporters. The foundation can evolve.

Performance claims come with caveats. In benchmarks released by the company in late July, Thomson competed closely with frontier systems. It matched or exceeded Claude Opus 4.8 in some legal reasoning tasks. It outperformed GPT-5.5, Claude Sonnet 5 and Gemini 3.1 Pro across a mix of evaluations. The tests covered instruction following, long context, coding and professional workflows.

Yet the comparisons aren't apples to apples. Thomson benefited from test-time scaling and internal retrieval tools linked to Westlaw and Practical Law. Competing models searched the open web in some evaluations. Internal composites formed part of the mix. Still, the results impressed enough academics. One preferred Thomson's responses even when others answered correctly. Citation quality held up against the leaders.

The first real test comes inside CoCounsel Legal. Thomson becomes the default for Tabular Analysis. That feature reviews up to 10,000 documents and fields as many as 100 questions about them. High-volume. Structured. Measurable accuracy. Exactly where a purpose-built model should shine. The broader CoCounsel platform, refreshed on August 20, now incorporates Anthropic's Claude Agent SDK for agentic workflows. It plans, reasons and executes across tools.

Hron doesn't hide the continued partnership. Thomson Reuters expanded its deal with Anthropic in May. "Our main objective is to make Thomson the model that powers more and more of CoCounsel's capabilities over time," he said. The in-house system supplements rather than supplants. For now.

This hybrid strategy reflects hard realities in professional services AI. General models hallucinate. They sycophantically please users. They lack deep context in regulated fields. Thomson trains explicitly to flag uncertainty. It avoids forcing confident answers when data runs thin. That choice reduces certain failure modes critical for lawyers and tax professionals.

Retrieval remains essential. The model pulls from verified sources. Outputs link back to Westlaw, Practical Law or Checkpoint where possible. Thomson Reuters stops short of promising line-by-line traceability for every claim. "We wouldn't claim every output can be traced line by line to an exact statute or ruling," the company stated. Professionals still verify. The system simply makes that verification easier.

Training on proprietary data changes the control equation. The company shapes tradeoffs between helpfulness and accuracy. It optimizes for professional caution over casual conversation. And it avoids feeding customer data to third-party labs, maintaining strict contractual prohibitions.

Recent coverage highlights the nuance. The New Stack noted the $40 million investment focused on post-training and expert evaluation rather than pre-training from scratch. It also detailed how Thomson integrates with agentic systems while preserving retrieval-augmented generation. The model doesn't replace retrieval. It enhances it.

The Next Web added color on the Alibaba connection and academic feedback. It quoted researchers who found Thomson's outputs preferable in quality and citation strength. The article also flagged Alibaba's recent moves toward charging heavy users of Qwen, underscoring why owning adaptations matters.

Thomson Reuters published its own early benchmarking in July. CTO Joel Hron and Head of AI Research Jonathan Schwarz wrote that capable models no longer emerge solely from frontier labs. One now comes from their own organization. The piece emphasized combination: authoritative content, expert judgment, professional tools and model development.

A smaller open-weight version of the model sits on Hugging Face for researchers. Commercialization beyond Thomson Reuters products remains under consideration. Customers won't buy direct access to Thomson-1 today. It lives inside the company's ecosystem, starting with legal research and document tasks.

The timing feels deliberate. CoCounsel has scaled to serve hundreds of thousands of professionals. Demand for trustworthy AI in legal and tax grows as regulators tighten rules. Enterprises want options that reduce vendor lock-in and API costs that can run high at scale.

Critics will note the $40 million figure pales against the billions poured into frontier labs. Success here hinges on whether domain-specific tuning plus retrieval consistently beats general models on real workflows. Early signs look promising. But benchmarks tell only part of the story. Real adoption will come from measurable productivity gains and risk reduction for law firms and corporate departments.

Thomson Reuters isn't alone in this push. Other data-rich incumbents eye similar paths. The difference lies in execution. Decades of curated content. Teams of domain experts. A product portfolio already embedded in professional routines. Those assets turn training data into a durable advantage.

Hron's house-buying metaphor lingers. Renting powerful models delivers immediate capability. Owning one, even if built on someone else's foundation, creates long-term optionality. The company can iterate faster on its own data. It can tune for fiduciary-grade standards. It can expand across tax, compliance and news without renegotiating every capability.

CoCounsel's latest agentic features show the complementary play. Built on Claude's SDK, the system orchestrates complex legal tasks. Thomson handles the heavy lifting on document volume and structured analysis. The combination aims for something greater than either alone.

Questions remain about geopolitical angles. Reliance on a Chinese open-source base, even heavily adapted, invites scrutiny in some markets. Thomson Reuters stresses the realignment process and independence going forward. Nothing locks them to Qwen. Future versions could draw from other bases or further internal development.

For industry watchers, this launch marks a maturation point. Pure reliance on API calls to Anthropic, OpenAI or Google gives way to selective ownership. The $40 million bet tests whether incumbents with rich datasets can close the gap on frontier labs in narrow but valuable domains.

Results so far suggest the answer leans yes for certain tasks. Thomson tops some composites on legal hardness and long-context handling. It admits when it doesn't know. It cites sources. These traits matter more to a partner at a law firm than raw benchmark scores.

The road ahead involves wider rollout. More features in CoCounsel. Potential expansion to tax and regulatory products. Continued benchmarking transparency. And ongoing collaboration with Anthropic even as internal capabilities grow.

Thomson Reuters has placed its chips. The house it builds won't replace every rented roof. But over time that equity could compound into a meaningful lead in professional AI. Lawyers and compliance officers will decide if the bet pays off. Their verdict will shape the next wave of enterprise AI strategy.