The Vertical Model Coup

Frontier labs are losing their grip.

Thomson Reuters just released benchmark data for their proprietary model, Thomson.

It beats GPT-5.5. It beats Gemini 3.1 Pro.

They did it with less than 10% of their proprietary data.

General-purpose models are now second-tier tools for the professional elite.

The Data Moat is Real

For two years, the tech press told you that scale was everything.

They said the company with the most GPUs wins.

They were wrong.

Thomson Reuters took an open-source foundation and layered it with decades of Westlaw, Practical Law, and Reuters data.

They used expert judgment to validate the reasoning.

They did not just scrape the internet for garbage text.

They used the gold standard of legal and professional information.

The result is a model that does not hallucinate like a bored intern.

It reasons like a partner at a Magic Circle law firm.

The Failure of General Intelligence

General models are great at writing poems about toast.

They are terrible at the high-stakes nuances of regulatory compliance or complex litigation.

When the cost of being wrong is a multi-million dollar fine, "mostly right" is a failure.

Thomson Reuters is proving that verticality is the only way forward for the enterprise.

They are not renting intelligence from Sam Altman.

They are weaponizing their own history.

This is a power move that every CEO should study.

If you are feeding your proprietary data into a general model, you are training your future competitor.

If you are waiting for a frontier lab to understand your specific business logic, you will be waiting forever.

The Behind-Closed-Doors Reality

Inside the boardrooms of the Fortune 500, the conversation has shifted.

It is no longer about which chatbot is the funniest.

It is about who owns the weights and who owns the data.

Thomson Reuters is showing that you can build a world-class model without a billion-dollar compute budget.

You just need the right data and the right experts.

• Data Sovereignty: Own the inputs, own the outputs.
• Vertical Logic: General reasoning fails at the edge cases of law and medicine.
• Compute Efficiency: Smaller, specialized models cost less to run and perform better.
• Expert Validation: Human-in-the-loop is not a buzzword: it is a requirement for accuracy.

What this means for leaders

1. Stop buying generic tokens for specialized tasks.
2. Audit your proprietary data silos today: they are your only real defense against commoditization.
3. Fire the consultants selling AI transformation without a specific data strategy.
4. Build your own vertical stack or prepare to pay a premium to those who do.
5. Stop worrying about AGI and start worrying about the competitor who just fine-tuned a model on your industry's secrets.