Cost Series 01 — The Bill Is Not the Token Count

Working notes for a planned episode — an outline, not an article. Details will shift when it gets scripted. Series index: The Bill.

The thesis episode: why "minimize tokens" optimizes the wrong variable, and what to optimize instead.

The hook

Everyone's first instinct for a big LLM bill is "use fewer tokens." That instinct is now wrong — or at least, it's aiming at the cheapest thing on the invoice. Caching and model tiering quietly broke the rule that tokens ≈ cost, and once you see the real cost function you'll sometimes send more tokens to pay less.

What the viewer walks away with

• Tokens ≠ cost anymore: a cached read is ~0.1×, output is ~5× and uncacheable, and a frontier model is ~15–20× a small one
• Token-counting optimizes the cheapest axis (cached input) while ignoring the two expensive ones (output, model tier)
• The counterintuitive move: a big stable cached prefix can be cheaper than a lean prompt re-sent fresh each turn
• The right objective is cost per unit of useful work, not tokens per call

Beats

• Open with a "trim the prompt" optimization that saves tokens and barely moves the bill — then show why
• Draw the real cost function: fresh ×1, cache write ×1.25, cache read ×0.1, output ×5, times the model tier
• The reversal: deliberately add a cached prefix, watch cost drop — tokens up, bill down
• Name the three real levers (routing, cache, output) — each gets its own episode
• Close: "stop counting tokens; price the work"

Demo ideas

• Same task two ways: lean-prompt-no-cache vs fat-cached-prefix — put both bills on screen
• A one-slide teardown of a real request's cost by axis (fresh/write/read/output)

Source material

• The series index — the cost-model reference numbers
• The token-usage / cached-read analysis — real invoices for the teardown
• Fleet Series 09 (the router) — the "cost acceleration" seed this series expands

Open questions

• How much pricing detail to put on screen vs. keep directional (prices change; the shape is durable)
• Lead with the caching reversal or the model-tier lever? (Caching is the more surprising hook)