You can’t ask for cheaper

Claim: telling the model to save tokens does not make it work smarter — it makes it skip the parts that cost tokens, which are the tests, the reviews, the reading-before-editing, and the comments and docs it would otherwise have written. Those are exactly the parts that keep the work correct and legible. You save on the bill and pay it back in bugs. Real economy is engineered on purpose; it cannot be requested.

The incident

I tried it the direct way once: I told the sessions to use fewer helper agents, to hold the cost down. Within a run the code started coming back unverified — tests thinning out, reviews skipped, changes landing without the checks I had not realized those extra agents were carrying. I put the limit straight back and gave up on the cheaper-model idea as well. The frugality had come out of the one place it must never come from: the checking.

The image

Ask a restaurant to make it cheaper and you do not get a cleverer kitchen — you get yesterday’s leftovers. You get what you pay for. The only honest way to lower the bill is to strike something you did not need: “take the water off, I drank some before I came in.” You cannot economize on the meal itself. A model told to be frugal does what the kitchen does — it hands back the work with the costly parts left out, and the costly parts were the checks.

The pattern

The instinct is natural: the sessions are expensive, so ask the agent to be frugal. But the tokens an agent spends are mostly spent checking — running the test suite, reviewing a change, reading a file before editing it, confirming its own claim. Ask it to spend fewer, and those go first, because they are the least visibly “productive.” The output looks lean and the defects show up later.

Comments and docs go the same way, and they are doubly exposed: nothing rewards them in the first place (the docs aren’t on the test), so “use less” trims what was already the thinnest. And the model cannot tell “trim your chatter” — which is safe — from “drop the comments and docs in the deliverable” — which is not; a blanket instruction cuts both.

There is a deeper reason it can’t simply be asked for: the model has no running meter. It does not size its own work by cost — it will fan out as many helpers as a task’s shape suggests, each on the most expensive model, without ever counting (agents launch at full price). “Use less” lands on a system with no sense of how much it is using.

The rule

Cost is cut by design, not by instruction. Real savings are deliberate and structural — send the cheap, mechanical steps to a cheaper model, narrow the scope so there is less to do, do the work once and reuse it — and each is a decision someone makes and enforces, never the agent choosing to try harder at frugality. The concrete levers are their own note (where the savings are): they all come down to taking off the bill something you did not need or already had.

The one budget you never raid is the checking. The tests and the reviews are the gate (a gate you can fail) — the only thing that can turn bad work away before it ships. Cut the gate to save tokens and the saving comes straight back as the bug it would have caught.

When cheaper is fine

There is a real exception, and it is the same principle from the other side: you can cut quality on the dimensions you genuinely do not need. A throwaway script to do X once, a small mock you want to poke at and throw away — no docs, no comments, no hardening, nothing kept or extended. Here “cheaper” is honest, because you are not raiding a gate you need; you are declining quality you were never going to use.

The tell is that you can name the one thing that must hold, and nothing else. You are renting the cheapest car in the lot just to try it: it only has to start and not explode, so you drive it a hundred meters and hand back the keys. You are alone in it and the trunk is empty. Its comfort, its longevity, its resale value come off the bill, correctly — you are not buying it, keeping it, or carrying anything in it.

The edge to watch: this holds only while nothing real rides in it, and two things end that. One is time — throwaway prototypes have a way of getting kept and extended, and the moment you keep one, every quality you cut turns into debt (the face transplant). The other is load — the moment your family is in the car or your suitcases are in the trunk, it is not a test drive, it is transport. Now a crash hurts someone, and even a clean trip gets your luggage covered in dirt. Real passengers and real cargo need the quality you were cutting. Ask for cheaper when you mean it; be honest about when you stop meaning it.

Prior art

Verdict: PARTIAL — the pieces are documented, the composite isn’t. Concise-chain-of-thought work shows brevity trading against correctness on hard problems (Renze & Guven, arXiv:2401.05618: a ~28% accuracy hit on math under “be concise”); TALE (arXiv:2412.18547) shows cutting tokens safely needs an engineered budget, not a vague ask; TRIAGE (arXiv:2605.13414) and BAGEN (arXiv:2606.00198) find models poor at sizing their own compute — the closest thing to “no running cost meter.” A JetBrains “talk like a caveman” benchmark (2026) found terse prompting trimmed only chatter with no quality loss — a useful boundary: cutting prose is safe, which sharpens that it is the verification-cutting that hurts. The delta: tying “ask for cheaper” to the specific failure of dropping tests, review, and reading, and “no cost meter, so economy must be engineered not requested” as one stated principle.