Cheap should mean stricter, not looser

There is a trade everyone in this industry believes in. On one side, frontier models: expensive, but you can trust the output. On the other, smaller and cheaper systems: affordable, but you accept that quality will slip. Budget-constrained buyers are told, politely, that they are choosing the second-best version of the thing.

We no longer think that trade is real. Or rather: it is real, but only for organisations that cannot say precisely what they want. Vagueness is the thing that costs money.

What you are actually paying a large model to do

When a system hands a broadly-specified task to a very large model, most of the capability being purchased is not domain expertise. It is inference under ambiguity. The model is being asked to work out what was meant, what the standard is, what a good answer would look like in this context — and then to produce one. That reconstruction is expensive, it happens on every single call, and it is thrown away the moment the response is returned.

In education this is especially visible, because the standards are not actually mysterious. What makes a lesson well-sequenced, an explanation age-appropriate, an assessment item diagnostic rather than decorative — these are known things. They are the accumulated content of a professional discipline. Every time a system asks a general-purpose model to intuit them from scratch, it is paying frontier prices to approximate knowledge that pedagogy has already written down.

Precision is what makes small models sufficient

The route to lower cost, then, is not a cheaper model asked to do the same vague job. It is a stricter specification of the job. Once you can state exactly what good looks like — as an explicit rule, a checkable constraint, a deterministic gate — you no longer need a model to infer it. You need a model to do the narrow creative work that remains, and that work turns out to be far smaller than the industry's default architecture assumes.

This is the part that surprises people: the effect on quality is not neutral, it is positive. An explicit standard applies identically every time. A model inferring the standard applies it differently on Tuesday than it did on Monday, and neither run is auditable. Moving a requirement out of the prompt and into the system makes it cheaper and makes it more consistent — the two supposedly opposed variables move in the same direction.

Why the industry frames it as a trade

Partly incentives: the economics of the frontier depend on capability being the answer to every question. But mostly it is that specification is unglamorous work. Writing down what good teaching looks like, precisely enough to be machine-checkable, is slow, domain-heavy, and impossible to shortcut with a bigger cluster. Buying scale is a way of deferring that work indefinitely, and it is a perfectly rational deferral right up until you have to run at national scale, on a national budget, and defend the output to a ministry.

Rigour is the compression. Everything you can state exactly, you no longer have to pay a model to guess.

So the cost-sensitive deployment and the high-standards deployment are not opposite ends of a spectrum — they are the same engineering programme viewed from two directions. The organisations that will serve the classrooms with the least money are not the ones willing to accept the loosest standards. They are the ones who were disciplined enough to define their standards well enough to enforce them in code.

Cheap, done properly, is stricter. That is not a consolation for the constrained buyer. It is the whole argument.