Here We Go Again: Thomson Reuters Is Training Its Own Legal AI Model

"Thomson" is the name Thomson Reuters gave to its new, proprietary legal LLM, launching this summer. If you feel like we're living in Groundhog Day, you're right here with me. Wasn't this the same story we heard in 2023 after the GPT-3.5 launch? Haven't we all come to the conclusion that it doesn't work? And yet, here we go again.

The real threat isn't Anthropic launching Claude for Legal

CTO Joel Hron did the press rounds in late March, including a solid interview with Artificial Lawyer. Asked directly whether Thomson is a hedge against OpenAI or Anthropic going vertical into legal, he said no, he doesn't think they want to do that, and TR will keep working with them. Thomson is "additive, not defensive," meant to complement the frontier models that already power CoCounsel.

Fair enough. Foundation model companies launching their own legal vertical would need to build trust in one of the most conservative professions on earth and compete against incumbents who've been doing this for decades. Probably not happening anytime soon.

But that was never the real threat anyway. The threat is a 40-lawyer firm figuring out that Claude Pro already does 80% of what CoCounsel does (in some cases more), for $20 a month per user. CoCounsel's pricing tells this story on its own. The entry-level product, CoCounsel Essentials (a thin Claude Sonnet wrapper for drafting and redlining inside Microsoft Word, no Westlaw research included) runs at $360 USD per month for a single user.

Want actual legal research grounded in Westlaw? That's $784 USD per month for a seat. Sure, there are (generous) bulk discounts for CoCounsel as enterprises purchase more seats, and a raw Claude Pro subscription has aggressive rate limits that make it impractical for deadline-driven legal work. But purpose-built legal AI tools like Sonar Legal are already shipping contract review and research workflows on top of frontier models at a fraction of CoCounsel's cost, and without the Claude rate limits. That gap is the kind of number that makes legal innovation departments ask uncomfortable questions.

Thomson as marketing, Thomson as moat

Let's dig deeper into economics. TR is a dominant player in Legal AI, with over a million CoCounsel users, and a brand that allows them to charge their services at a significant premium. But that premium needs justifying, and Thomson (the model, the brand, the marketing narrative) is how they plan to do it.

On the cost side, owning their own model gives TR more control over inference economics. Every CoCounsel query that currently hits an Anthropic or OpenAI API costs TR real money per token, and that cost goes up as workflows become more agentic (we're talking 10 to 30x more tokens per task compared to a simple chatbot query). Owning the model converts variable API spend into fixed infrastructure cost. On its face, that's smart treasury management.

But the more interesting play is the narrative. "Our AI is specifically trained for legal. The general models are general. You need specialized intelligence, built on 175 years of editorial expertise, validated by 4,500 subject matter experts." That story is aimed squarely at the committee weighing whether to renew a CoCounsel subscription or just give everyone a ChatGPT Business license and call it a day. Thomson is TR's answer to the question "why should we pay you ten times what the foundation model costs."

Whether that answer holds up depends entirely on Thomson's output quality, as well as its ability to keep up with foundation models.

Why Thomson will not be better than foundation models

How good will Thomson actually be? If history is any guide: not good enough. Every attempt at a domain-specific legal LLM has run into the same wall. You spend months curating data, fine-tuning, validating, benchmarking, and by the time you ship, the next generation of foundation models has already leapfrogged your carefully tuned system on the very tasks you optimized for. It happened in 2023, it happened in 2024, and there's no reason to think 2026 will be different. The general models keep absorbing what used to be specialized capabilities, and they do it on a timeline that no single company's retraining cycle can match.

Has TR actually managed a way to tackle this issue? Hron made a claim in the Artificial Lawyer interview that's worth unpacking. He described Thomson as "portable," the idea being that when a better open-source base model comes along, TR can just move Thomson's legal training over to it. They're never stuck with one foundation.

That framing is generous. What's actually portable is TR's process: the curated legal datasets, the benchmarks, the recipes for how they train. What's not portable is the trained model itself. Fine-tuned weights are tied to the specific base model they were built on. When a meaningfully different open-source model drops (and they drop constantly), TR doesn't migrate. It retrains. That means months of work, not to mention the additional costs: re-running the legal fine-tuning, re-validating across benchmarks, re-deploying. And while TR is doing that, every competitor running on APIs has already switched to the new frontier model, often within days.

Hron says Thomson already outperforms general models on four of ten key legal benchmarks. That sounds promising until you consider that the general models he benchmarked against are already a generation old. GPT-5.4 shipped in March. Claude Sonnet 4.6 landed in late February. By the time Thomson launches this summer, the targets will have moved again.

So when a law firm paying $784 a month per seat asks whether TR's specialized model actually produces better work product than the latest Claude with good prompting and their own documents, the answer, based on every previous attempt to out-specialize a frontier model, will almost certainly be no. Thomson's retraining cycle will never be fast enough. The benchmarks TR is winning today were run against models from months ago, and months in this market is a lifetime.

The broader economics of Legal AI

TR building Thomson is the strongest signal yet that the economics of legal AI's current model are strained. A company with dominant market share, a billion-dollar acquisition behind its AI product, and hundreds of millions in annual AI investment has concluded that depending on third-party APIs for its core product is a risk worth spending heavily to mitigate. That tells you something about where the margins are, and where they're going.

But the lesson for the rest of the market isn't "go train your own model." Most legal AI companies can't, and the ones that try will find themselves perpetually a few months behind the frontier, spending engineering resources on model infrastructure instead of product development. The lesson is that the value in legal AI increasingly lives in the workflow layer (retrieval architecture, domain logic, quality controls, feedback loops, the ability to integrate with how lawyers actually work) and that this layer works best when it sits on top of the most capable model available at any given moment. If you're locked into a model you retrained six months ago, you're making a choice to be worse at the foundation layer in exchange for better control over costs and narrative. That's a trade-off that makes sense when you're Thomson Reuters protecting a franchise. It makes considerably less sense for everyone else.

The firms and legal departments that figure this out, that treat the model as a commodity and the workflow as the product, are the ones that will be hardest to displace. By TR, by anyone.