# Right-size the model

There is no "an AI." There are dozens of models, and in one controlled test the
price ran **143× apart for the identical job** — A$0.52 to A$73.00 to build the
same working app. The decisive finding: **price did not predict quality.** A 93¢
model reviewed code better than the $73 one. Picking the right model per task is
a real, measurable cost discipline.

## The rule

> **Default to cheap. Pay up only where a wrong answer is genuinely expensive.**

## A routing cheat sheet

| Use the cheap/fast model for… | Use the frontier model for… |
|-------------------------------|-----------------------------|
| Boilerplate, scaffolding, glue code | Subtle correctness on money / safety / security |
| Drafting, summarising, reformatting | Tricky concurrency, algorithms, data integrity |
| High-volume, repetitive calls | Architecture and design trade-offs |
| First-pass review (then verify) | The final sign-off on something irreversible |
| Anything you'll verify anyway | One-shot work you can't easily check |

## Concrete tiers (re-check every quarter — today's cheap model was last year's frontier)

| Tier | Reach for it when | Rough cost shape |
|------|-------------------|------------------|
| **Economy** | Default. High volume, low blast-radius, you'll verify anyway. | cents per task |
| **Mid** | Real reasoning, but a wrong answer is cheap to catch and fix. | low single dollars |
| **Frontier** | Irreversible, hard-to-detect, or safety/money-critical correctness. | $10s per task |

Anchor those tiers to your own measured numbers, not a vendor's list price.

## How to think about it

1. **What does a wrong answer cost here?** Cheap to catch and fix → economy.
   Expensive, irreversible, or hard to detect → pay up.
2. **Will I verify it anyway?** If yes, let the cheap model draft and spend your
   judgement on the check — not on the generation.
3. **Volume changes the maths.** At 1 call the price gap is noise. At 10,000
   calls a 100× gap is your whole budget. Right-sizing is a budget control, not a
   preference.

## Don't forget

- The biggest cost is usually a *wrong answer shipped*, not the API bill. Route
  the model to match the cost of being wrong — then verify either way.
- For a worked example — same brief, eight models, full cost and quality read —
  see the experiment at **kaikalinowski.com/demos/compare/**.