⚖️ Machines That Talk · Day 15

Cheap, Fast, Smart: Pick Two

There is no single best AI. Big models are super smart but slow and pricey. Tiny models are cheap and fast but kind of dumb. The pro move isn't picking the biggest — it's matching the model to the job. Let's play matchmaker.

↓ start matching
🎛️ Lab · The Model Matchmaker

Pick a job. Hire a model.

Here are your four AI workers — from a tiny 🐣 to a giant 🦣. Each one shows how smart it is, how fast, its price, and how much text it can hold (its context). Pick a job, then tap a model to hire it — the real bill and the verdict appear live.

👆 Tap a model aboveHire one of your four workers for this job and I'll tell you if it's too weak, overkill, or a perfect fit — and exactly what it costs.

🧠 smart · ⚡ speed (out of 5)   💵 price = $ per 1,000,000 tokens (in / out)   📏 context = how much text it can read at once.

🐞 Break-It Challenge

Two ways to pick wrong

The two classic mistakes, side by side. Hire the cheapest worker for the hardest job and it flat-out fails. Hire the priciest worker for the easiest job and it works — while burning a shocking pile of money for nothing.

See the trap? A tiny model isn't just "a bit worse" — on the wrong job it returns nothing usable. And the giant model isn't just "a bit pricey" — it can cost hundreds of times more for a job a 🐣 could nail. Wrong model = failed job or wasted money.
🏷️ The pro wordChoosing the right model for a task is called model selection. When an app does it automatically — reading each request and sending it to the best-value model on its own — that's model routing. Every big AI product secretly routes your messages this way to stay cheap and fast.
🧭 Boss Build · Route It Right

Be the router

A job comes in. Your call: send it to the best-value model — the cheapest one that can actually handle it. Too weak fails. Too fancy wastes money. Only the smart-cheap pick counts. Route 3 jobs right to earn the final badge!

Routed: 0 / 3
🚀 Beyond the Basics

How the pros really route

Choosing a model by hand is where you start. Here's the ladder from "one model for everything" to how real AI products keep costs low at massive scale:

🦣

One model for everything (what you'd try first)

Just send every request to the biggest, smartest model. ✅ Dead simple, always capable. ⚠️ You pay giant-model prices even to say "hi" — costs explode.

🧭

Route easy vs hard (what you built)

Read each request and send it to the cheapest model that can handle it. ✅ Big savings, still smart when it matters. ⚠️ Guessing "how hard is this?" is its own tricky problem.

🪜

Cascade & distill (the real deal)

Try a cheap model first and only escalate to a big one if it struggles — or train a small model to imitate a big one. ✅ Best value at scale. ⚠️ Complex to build and tune.

🏷️ Pro names to look upAuto-picking per request is model routing. Cheap-first-then-escalate is a cascade. Training a small model to copy a big one is distillation. And running non-urgent jobs overnight for a discount uses the batch API — all tricks to get cheap, fast, and smart.

🎉 Day 15 Complete!

You learned the pro's secret: there's no single best AI. Smart, fast, and cheap pull against each other, so you match the model to the job — the cheapest worker that can actually do it. That's model selection, and doing it automatically is routing.

Day 16: Give the Bot a Notebook →
Tomorrow: our bot forgets everything the moment you close the chat. We give it a notebook so it can remember and look things up. 📓
AI Adventure · Machines That Talk · Day 15 — made for young inventors 🚀