Thomson Reuters built its own AI instead of renting one, for $40 million

At a moment when OpenAI and Anthropic are slashing prices to fend off Chinese AI rivals, Thomson Reuters chose to skip renting a frontier model entirely. Instead of taking advantage of cheaper APIs, it poured money into its own model — the logic being that for repetitive, high-volume work, running your own model ends up cheaper than paying someone else's subscription fees.

From Qwen to Snowdon to Thomson

Working with Imperial College London, Thomson Reuters first retrained Qwen for safety, ethics, and political neutrality, naming this intermediate model "Snowdon" after the Welsh mountain. From there, the company pretrained it further on its own proprietary content, ran additional fine-tuning with domain experts, and layered on agentic reinforcement learning inside its own tooling environment. Despite all these stages, the company says it has used less than 10% of its total content library so far.

CTO Joel Hron said the team has already swapped out the open-source starting point "close to five or six times," while head of research Jonathan Schwartz pointed out that the bigger achievement isn't any single model but the "model factory" the process has produced.

Benchmarks favor Thomson only on its own data

The company's blog post frames Thomson as a world-class model, but the numbers the company itself released paint a more cautious picture.

Benchmark	Thomson	Comparison
Stanford LegalBench	0.823	Trails Gemini 3.1 Pro and GPT-5.5
Harvey Legal Agent Benchmark	-	Ranks right behind Opus 4.8
In-house Deep Research (web access only)	0.53	GPT-5.4 scores 0.65
In-house Deep Research (with proprietary content)	0.83	GPT-5.4 scores 0.82

Thomson leads on PrBench Legal, a benchmark for instruction-following and tough legal questions, but falls well behind on reasoning and coding. The comparisons aren't entirely apples-to-apples, either — Thomson competes using more test-time compute, while GPT-5.5 ran without its reasoning mode enabled.

On web access alone, Thomson is "within range of other models but not yet a clear leader," evaluation lead Andrew Bean admitted. What's notable is that GPT-5.4 gets roughly the same performance boost when given access to the same proprietary content. That suggests data access itself — not just specialized training — is what's driving the performance gap. It's also worth noting the company didn't benchmark against more recent models.

Why stick with a custom model anyway

Why would Thomson Reuters build a model from scratch instead of simply fine-tuning a frontier model from OpenAI or Anthropic on legal data? The company points to three reasons.

The first is economics. Schwartz noted that standard fine-tuning "has a strong tendency to degrade general performance," and fine-tuning also locks you into that provider's inference costs and roadmap going forward. For high-volume work like document review, the math favors a smaller, cheaper, purpose-built model.

The second is data. Performance jumps sharply when the model is trained directly inside proprietary tools like Westlaw, and the company has no intention of handing that kind of access to any outside provider.

The third is compounding value. Hron described it as "the difference between renting a house and owning one." Every product update accumulates expert-reviewed content into the training data. With your own model, that becomes a long-term asset that keeps growing in value. Rely on someone else's model, and that value flows to the provider instead.

Hron said future AI competitiveness will come down to "how you orchestrate it, and knowing which intelligence is important enough to own."

How it's being deployed

At launch, Thomson powers the Tabular Analysis feature in CoCounsel Legal, the company's legal AI product — a task the company judged economical enough to run on a small, low-cost model. The product itself still runs on a multi-model architecture, so administrators can swap models in and out, and Thomson isn't meant to orchestrate everything — it's handling narrower tasks like citation checking. The company says customer data is not used for training.

A smaller version will be released on Hugging Face under a non-commercial license, with a technical report and developer portal to follow. Hron said early, non-binding conversations about direct licensing with law firms are already underway.

Editor's take

What makes this interesting isn't a race to build the smartest model — it's the question of who's actually earned the right to own one. Thomson Reuters can run this math because it has three things at once: proprietary data, hundreds of full-time domain experts, and a workflow that lets it objectively score results. Companies with that combination can now build competitive, specialized models for $40 million, at a moment when open-source communities are catching up to frontier labs within months. Companies without proprietary data or a way to score outcomes tend to end up with nothing but an endless maintenance bill if they try the same thing.

There's another notable wrinkle in the generational comparison. On web access alone, Thomson trails GPT-5.4 — but GPT-5.4 gets the same performance boost as Thomson once it's given access to proprietary content. That's not the power of specialized training; it's the power of data access itself. In effect, the company is admitting its edge comes not from "our model being special" but from "the data only we have." Domestic companies with proprietary data and expert-verification systems in fields like law, medicine, or finance could apply the same math. But companies chasing the "build our own model" trend without that data or a way to score results are likely to end up just piling up costs.

In the coming months, it's likely Thomson Reuters will move to a stronger Qwen variant and push past that 10% data-utilization ceiling. Whether the next version can beat frontier models on web-only access, without the proprietary-data boost, will be the real test of whether this experiment actually worked.