Explain like I'm five · AI

What is Pricing?

LLM pricing (per-token billing), told as a story about a parrot that is priced per bean, not per visit.

trainerMONTH-END BILLexpected: 100,000actual: 1,000,000 !…what cost so much?the parrot just answeredthe bill is ten times the guess — and nobody knows where the beans went

The month-end bill is ten times the guess. And nobody knows what cost so much.The parrot only answered as told. Nobody counted where the beans leaked away.

BILL LINESparrot rental, 1 time … nobeans in × ratebeans out × raterate by parrot sizepriced per beanone bean = one charge₩₩₩₩beans going in cost too₩₩₩₩₩₩beans coming out cost morethe parrot is priced per bean, not per visit — uncounted beans leak away

The parrot is priced per bean, not per visit.Beans going in and beans coming out both cost. What a bean is: the parrot counts words in beans.

cost = beans in × rate + beans out × ratebeans in × 1todayweatherhow?₩1 per beanthe bigger the parrot, the pricier per beansunnywarmtoday₩₩₩₩₩₩₩₩₩beans out × 5beans out cost 3 to 5 times morepricing: beans in × rate + beans out × rate — and a bigger parrot costs more per bean

Pricing is beans in × rate plus beans out × rate. Beans out usually cost 3 to 5 times more.And a bigger parrot costs more per bean. So use fewer beans, or a smaller parrot, or both.

Keep the tray shortfewer beans going in
Cap the beans coming outset a maximum length
Cache the same startbeans laid out ahead cost a tenth
Easy jobs to the small parrotten times apart per bean
MONTHLY BILL — BIG PARROTbeans in 1,000,000 × 3 = 3,000,000beans out 200,000 × 15 = 3,000,000cached beans 500,000 × 0.3 = 150,000total 6,150,000(a small parrot: about 620,000)big parrot3 / 15 per beansmall parrot0.3 / 1.5 per beansame beans, ten times apart by parrot sizeread the bill as bean kind × parrot size — it shows where the money goes

Read the bill as bean kind × parrot size, and it shows where the money goes.200,000 beans out cost the same as 1,000,000 beans in. Shrink the tray (context), reuse beans laid out ahead for the same start, and route easy jobs to the small parrot.

₩₩the price of one bean falls every yeargood newsthe muttering parrotone errand or think-aloud: thousands of beansbean counts grow ten, a hundred timesso the bill grows againbeans get cheaper, but you use more — counting is the answer

One bean gets cheaper. But bean counts grow ten, a hundred times.The errand-running parrot and the muttering parrot eat a great many beans. How cost and time add up: what an answer costs.

In one breath

Pricing = beans in × rate + beans out × rate. Beans out cost 3 to 5 times more, and a bigger parrot costs more per bean. Keep the tray short, cap the output, cache the same start, send easy jobs to the small parrot.

LLM APIs bill input tokens and output tokens at separate rates that vary by model tier (output is usually 3 to 5 times more). Prompt-cache hits earn a large discount, and batch APIs are the slow-but-cheap lane. Context length, output length, and model choice drive cost; agents and reasoning models use so many tokens that totals can rise even as rates fall.

When grown-ups say it

Token rate
The price of one bean. Usually listed per million beans. It differs by parrot size.
Input · output tokens
Beans in · beans out. Beans out cost 3 to 5 times more. Long answers add up fast.
Cache discount
Beans laid out ahead are cheap. Reusing the same start costs about a tenth. → beans laid out ahead
Context length cost
A bigger tray costs more. Every bean on the tray is charged every time. → the tray in front of the parrot
Model tier
Parrot size. A big parrot can cost ten times more per bean. Match it to the job.
Batch API
The slow but cheap lane. Gather questions that can wait, and they get answered at half price.
Budget cap
A lid on the bean sack. Only this much a month — set it to stop before it runs over.
Routing
Which parrot gets the job. Easy jobs to the small parrot, hard ones only to the big. → which parrot gets the job
Next story
The Parrot's Gaze →
Want another term as a picture book? Request →