Thomson Reuters announced on 24 August 2026 that it has built its own frontier-grade large language model, trained on decades of proprietary legal and tax content, for roughly $40 million. The model, called Thomson, is now deployed inside CoCounsel Legal, the company’s AI assistant for lawyers.
The claim is striking: a model that performs competitively with GPT 5.4 and Claude Sonnet 5 on legal tasks, built for less than 5 per cent of what frontier labs spend. The final training run cost $450,000.
What Thomson Actually Is
Thomson is not a general-purpose model built from scratch. Thomson Reuters started with an open-source foundation — most recently Qwen 3.5 — and specialised it through mid-training and post-training on proprietary content from Westlaw, Practical Law, Checkpoint, and Reuters news. Hundreds of the company’s own lawyers and tax professionals created training examples, evaluated outputs in blind head-to-head comparisons against frontier models, and judged whether citations actually supported the model’s claims.
The result, according to Thomson Reuters’ internal benchmarks, is a model that matches or exceeds frontier models when given access to its own content ecosystem, but trails them when working only with web-sourced information. That gap is the point: Thomson’s advantage comes from proprietary data and tool integration, not raw parameter count.
Crucially, Thomson Reuters says it has trained the model on less than 10 per cent of its content so far. The company frames the remaining 90 per cent not as more of the same, but as “continued discovery of new kinds of specialization.”
The $40 Million Question
Frontier labs have spent billions training their flagship models. Thomson Reuters spent $40 million over two years, covering both talent and compute. CTO Joel Hron’s framing: “For years, the AI industry has treated scale as the answer: bigger models, more compute, more money. Thomson shows there is another path.”
Whether that path generalises beyond legal and tax work is the open question. Thomson Reuters is explicit that Thomson “does not need to keep pace with the frontier of general intelligence across all dimensions” — it needs to set the frontier for legal. That is a narrower ambition, and a more achievable one.
The company is also making a “small” version of Thomson available as an open-weight model on Hugging Face under a non-commercial academic license, inviting outside researchers to test its claims.
AI Sovereignty as Product Strategy
The launch positions Thomson Reuters as more than an integrator of other companies’ models. CEO Steve Hasker’s framing: “The company has always owned the content, the expertise, and the tools professionals rely on every day. Now it owns the model too.”
This connects to a broader shift in how organisations think about AI sovereignty — who controls the training data, where the model runs, and what biases live inside it. Thomson Reuters is already in conversations with large law firms about direct licensing, including firms that want to fine-tune Thomson with their own internal data. Hron said Thomson was “built as infrastructure for Thomson Reuters” but could “also become infrastructure for others.”
This is the same logic driving sovereign AI initiatives across governments — the idea that relying entirely on third-party models creates strategic vulnerability. Thomson Reuters is applying it at the corporate level, betting that professionals with fiduciary duties will prefer a model they can audit, control, and run on their own terms.
What This Means for Professional Work
The model’s first deployment is narrow: Tabular Analysis in CoCounsel Legal, a structured document review tool. Thomson Reuters says CoCounsel will remain multi-model, using Thomson where it has an advantage and other models elsewhere.
But the trajectory is clear. Hron expects Thomson to take “a bigger and bigger share of the tokens” over time, and the company plans to extend it across its legal and tax portfolio. If Thomson’s claims hold up under external scrutiny, the implication is that domain-specific models — trained on high-quality proprietary data and evaluated by actual practitioners — can compete with frontier generalists at a fraction of the cost.
That matters beyond law. Accounting, compliance, medical diagnostics, engineering — any field with deep proprietary knowledge bases and high accuracy requirements is a candidate for the same approach. The era where only a handful of frontier labs could build capable AI may be narrowing.
The NZ Connection
New Zealand’s legal sector is small and concentrated, but the sovereign AI question is directly relevant. Large law firms and government agencies here rely on overseas models — primarily OpenAI and Anthropic — for AI-assisted legal research and document review. Thomson Reuters’ model factory approach suggests an alternative: start with a strong open-source base, specialise on local legal content, and keep control of the result.
Whether NZ has the content depth and expertise to replicate this is debatable. But the broader point — that capable AI is becoming cheaper to build and easier to control — aligns with the sovereign AI arguments we’ve explored in NZ’s sovereign AI moment and the APAC sovereign AI race. Domain-specific models may be where smaller players can actually compete, rather than chasing general intelligence they cannot match.
This also echoes the competitive dynamics we’ve seen in legal AI specifically, where companies like Legora reached a $5 billion valuation by building specialist tools rather than generalist models. Thomson Reuters is taking that logic to its extreme: owning the model itself.
