One of the Biggest Legal Publishers in the World Just Built a Model. Look at What It Started From.

Thomson Reuters has launched its own frontier model. The interesting part is not that they did it. It is that they began with an open-source foundation, spent about 450,000 dollars on the training run, and still run other models alongside it. That is the whole argument about legal AI, settled by the company with the most to gain from the opposite answer.

The layer worth owning in legal AI is the verified content and the discipline that keeps it correct, not the model.

I have been making that argument for a year. In July I wrote that you can rent a model but you cannot rent case law discipline. Then in August, Thomson Reuters did the thing that ought to disprove it, and the detail of how they did it makes the case better than I could.

What they actually built

Thomson Reuters launched Thomson, its first proprietary large language model, in late August 2026.

The numbers they have published are worth reading carefully:

• About 40 million dollars over two years, on people and computing.
• A final training run of about 450,000 dollars.
• Built on an open-source foundation, with their own content, training methods and expertise added.
• Trained on less than 10% of Thomson Reuters content so far.
• Shaped by hundreds of internal subject matter experts.
• First deployment will be inside Tabular Analysis in CoCounsel Legal, which is high-volume structured document review. It will be the default model there, and administrators can switch it for another.
• CoCounsel Legal stays multi-model by design, using Thomson where it wins and other models elsewhere.

"Early evaluations put Thomson on par with the latest frontier models across a range of tasks." That is Thomson Reuters on Thomson Reuters' own model, and the benchmarks behind it are self-reported with no third-party verification.

Bob Ambrogi went through them at LawSites. He found the picture more mixed than the headline. Thomson took the top score in three of seven categories, lost the best-known legal benchmark, Stanford's LegalBench, to two rivals, came last on coding, and was tested under conditions that gave it test-time scaling while one competitor ran in non-reasoning mode.

But the finding that matters here is the second test. Thomson Reuters compared Thomson with Westlaw access against frontier models searching the open web, and won. Ambrogi's response was that this measures the retrieval source rather than the model, and would hold true whichever model sat behind the Westlaw connection. Pressed on it, Thomson Reuters disclosed that they had run the fair version: frontier models given the same Thomson Reuters content, on a simpler harness. The scores came in between 0.81 and 0.91. Thomson scored 0.89.

Read that again. On their own numbers, at least one general model with access to their content scored above their purpose-built model with access to their content.

The content did the work.

The bit everybody is skipping

Read those first three points again.

One of the two largest legal information businesses in the world, with Westlaw, Practical Law, Checkpoint and Reuters behind it, and a genuine strategic reason to own a model outright, did not build one from scratch. They took an existing open-source model as the foundation and added what only they have.

And the expensive part was not the model. It was two years and a lot of very expensive lawyers and engineers curating, structuring and verifying content. The actual training run, the bit that sounds like the hard part, came in at 450,000 dollars. That is less than a mid-sized firm spends on its practice management system in a year.

And they did not do it once. Thomson Reuters swapped out its open-source starting point close to half a dozen times over the life of the project, as better open models appeared, with the most recent run built on Qwen 3.5. Their head of foundational research put it plainly: the bigger finding is less the individual model than the model factory they built around it. They designed for the base model being replaceable, because it is.

That is not a story about model-building being newly accessible. It is a story about where the value sits. The foundation was commodity enough to start from, and swap, and swap again. The content and the expertise were not.

One caveat on the numbers. The 40 million dollars covers talent and computing since the work began. The team behind it arrived through Thomson Reuters' 2024 acquisition of Safe Sign Technologies, for an undisclosed sum that sits outside that figure. The model was cheaper than a frontier lab's. It was not 40 million dollars cheap.

And the buyers are asking for the same thing

That is the supply side. The demand side landed this week, from Thomson Reuters' biggest competitor.

LexisNexis published UK research on 2 September, surveying 543 legal professionals. 81% said they feel more comfortable using AI when it is grounded in legal sources, up from 72% in January and 70% a year ago. The single biggest concern in the profession is accuracy: 83% worry about lawyers relying on inaccurate or fabricated information, well ahead of data leakage at 53% and over-reliance at 52%. More than three quarters, 77%, think AI is causing clients or the public to misinterpret the law.

Adoption is not the question any more. 94% use AI for legal work and 74% use it at least weekly. What has moved is what they want from it.

Read that alongside the Thomson Reuters launch and you have both halves of the same argument, made by two competitors who between them own most of the legal content in the English-speaking world.

Be clear-eyed about it. Both are vendors. LexisNexis sells AI grounded in legal sources, so a survey finding that lawyers want AI grounded in legal sources is a survey finding a vendor was always going to be pleased with, and it is their own press release about their own research. Thomson Reuters' benchmarks are self-reported too.

But they are competitors, they approached it from opposite ends, and they arrived at the same place. One spent 40 million dollars proving the content is what moves the numbers. The other asked 543 lawyers and found four in five now want exactly that. When two rivals with opposing commercial interests agree on the shape of a thing, that is worth more than either saying it alone.

What this does to my July argument

It qualifies it, and I would rather say so than pretend otherwise.

In July I argued that only the firm layer, your own judgement and discipline, holds durable value, because the model underneath is rented and interchangeable. Thomson Reuters has demonstrated something I did not give enough weight to: the content layer holds durable value too, if you own enough of it.

They own one of the two largest verified bodies of legal content in the profession. That is a genuine moat, and pushing it into a model they control is a rational thing to do with it. My July piece was right that you cannot rent discipline. It was too quick to imply the content layer was equally rentable. For a publisher at that scale, it is not.

But notice what follows from that, because it cuts the other way for everybody else.

Why "own the content" is not advice a firm can act on

The moat is the size of the corpus, and almost nobody has one.

Thomson Reuters can justify 40 million dollars because they are amortising it across Westlaw, Practical Law, Checkpoint and Reuters, and across every customer of all four. A claimant firm has case files, precedents and a house style. Valuable, genuinely, but not a corpus you train a frontier model on, and not one you would want to, because the moment the law moves your model is wrong in a way that is expensive to fix.

That is the practical difference between a model and a knowledge base, and it is the whole ballgame for a firm.

A knowledge base can be corrected on a Tuesday afternoon. A judgment lands, somebody reads it, the entry changes, and every answer from that moment reflects the new position. A model cannot. Changing what a model knows means retraining it. Thomson Reuters is running theirs alongside other models rather than instead of them, and has used less than a tenth of its own content so far. Their stated reasons are about applying Thomson where it has the clearest advantage and continuing to discover new kinds of specialisation, rather than about correction speed. But the effect is the same. They are moving carefully, and they have every resource.

So the honest read of the last fortnight is this. Thomson Reuters has proved that content is the moat, which is what I have been arguing. They have also proved that turning content into an owned model is a publisher's move, not a practitioner's, and they have priced it for everyone to see.

What a specialist firm should own instead

Four things, in order of how much they matter.

The verified position on the law you actually work in. Not all law. The narrow band you litigate week after week, held as a knowledge base outside the model, with a named person responsible for it being current.

The discipline that keeps it current. A route from a new judgment to a changed answer, with a time attached. If nobody can tell you how long that takes today, that is the finding.

The check at the point of output. Every authority cited traced back to the entry it came from, so a fabricated citation cannot survive to the document. This is the step that most general tools skip, because it slows things down.

Your own house standard. How your firm words a chronology, structures a statement, handles a weak point. That is the layer nobody outside the firm can supply, and it is the one clients actually notice.

None of that needs a proprietary model. All of it needs somebody to own it.

How to tell whether a legal AI product has real discipline behind it

Five questions. They take a supplier about ten minutes to answer, and the answers separate the market quickly.

1. Where does your legal content come from, and who verifies it? A named process, not "we use leading models".
2. What happens when a judgment changes the position? Ask for the mechanism and the time from judgment to product.
3. Show me the last time it happened. If nothing has changed recently in an area that has moved, that is the answer.
4. Can I trace any cited authority back to your source? If a citation appears in the output, you should be able to click through to the entry it came from.
5. Which model is underneath, and what happens when you change it? A supplier who cannot switch models has a dependency. A supplier who can, and whose verified content survives the switch, has built it the right way round.

That last one is the Thomson Reuters lesson in a single question. They can swap models and their content still holds. That is what good looks like at any scale.

Where this leaves us

We built LegalDocs Assist on the argument that a general model plus a verified, curated body of credit hire law beats a general model on its own. The last fortnight has tested that from the strongest possible direction and it has come out better than it went in.

The company best placed to prove that owning a model is the answer has instead shown that the model is the cheap part, the content is the expensive part, and even they are keeping their options open underneath.

If you want to see what a verified knowledge base looks like in practice on credit hire, have a look at how it works, or get in touch and we will walk you through the sources.

© LegalDocs Assist — www.legaldocs-assist.co.uk