Frontier or open-weight? Decide per workflow, not per company
The choice between a hosted frontier model and an open-weight model you run yourself is an engineering decision. Making it once, for the whole business, is how you end up wrong twice.
By Stanley Watterson
Businesses tend to arrive with a position already formed. Either everything goes to a hosted frontier model because it is the best available, or nothing does because the data is sensitive. Both are the same mistake in opposite directions: a company-wide answer to a per-workflow question.
What actually decides it
Data sensitivity. Not whether the data feels private, but whether you have a contractual or regulatory obligation about where it goes. That is a question with a real answer, and it is worth getting it properly rather than guessing.
Cost per call against volume. A frontier model is cheap for a hundred calls a day and expensive for a hundred thousand. Open-weight models flip that: real setup cost, then marginal cost close to nothing. The crossover point is calculable.
How hard the reasoning is. Classifying an email by urgency is not a difficult task and a small open-weight model handles it comfortably. Reading a supplier contract and flagging unusual terms is genuinely hard, and the gap between the best models and the rest is still wide.
Tolerance for change. Hosted models get deprecated and replaced on someone else's schedule. A model running on your own hardware does exactly what it did last year, which is either reassuring or limiting depending on the workflow.
The mixed answer is normal
A realistic setup routes high-volume classification to something small running locally, and sends the handful of genuinely difficult cases to a frontier model. Most of the cost sits in the first path, most of the value in the second.
What matters is that the decision is written down per workflow, with the reasoning attached. When the pricing changes (and it will) you re-run one decision rather than re-litigating your entire strategy.