Stack Basics · Module 4 of 5
Module 4 — What this all costs, and the cost hacks that actually work
Hosting is cheap — Modules 1–2 run on about $6–12 a month plus $10 a year. The real money question in an AI stack is model usage: what you pay every time an AI writes code for you or powers a feature in your product. Here's how the costs actually behave, and the hacks people use — with the tradeoffs the hype posts skip.
Where the money goes: tokens
Every AI call is billed in tokens (roughly ¾ of a word each). You pay for what goes in (your prompt, your files) and what comes out (the answer). Coding agents are token-hungry — they read big chunks of your repo, think, retry. That's why the same model can cost pennies in a chat window and real dollars in an agent loop.
Two ways to pay:
- Subscriptions (Claude Pro, ChatGPT Plus, Cursor): flat ~$20/mo-ish for generous-but-capped usage. Best deal for individuals learning — your worst month is capped, and you'll use more than you'd dare to at per-token prices.
- API keys (pay-per-token): no cap, no floor. What you'll use inside your product (your server calling a model to serve a customer), and what tools like Aider run on. Costs scale with success — mostly a good problem.
The rough shape across every provider: frontier models cost 10–30× more per token than small models. That gap is the whole cost game.
Hack 1: model routing — the one that actually matters
Not every job needs the frontier model. Summarizing, classifying, formatting, drafting boilerplate — small cheap models do this fine. Architecture, gnarly debugging, code your income depends on — that's frontier-model work.
The practice: default cheap, escalate on failure. My own pipelines route almost everything to lower-cost models and only reach for a premium model when the cheap one demonstrably fails. That single habit routinely cuts model spend by 10× or more, with almost no quality loss — because most jobs weren't hard.
Hack 2: OpenRouter and free models — real, with a real tradeoff
OpenRouter gives you one API key for hundreds of models across providers, which alone is worth having: swap models by changing one string, compare prices in one dashboard.
It also lists genuinely free models — usually open-source models someone is hosting at zero cost, often rate-limited. Here's the honest version of the tradeoff:
- Fine for: learning, prototypes, personal scripts, high-volume/low-stakes jobs (tagging, drafts), anything where a mediocre answer costs you nothing.
- Not fine for: production code you'll sell, security-sensitive work, anything customer-facing. The quality drop is real — more subtle bugs, worse instruction-following, and free capacity can vanish or slow down without notice.
Rule of thumb: free models to learn and tinker, paid models where your name is on the output. A subtle bug in production costs more than the tokens you saved.
Hack 3: local LLMs — your hardware, zero per-token cost
Run open models on your own machine with Ollama (one install, then ollama run <model>). Truly free per token, fully private, works offline.
The catch is hardware physics:
- 8–16GB RAM: small models (7–8B class) — usable for chat, summaries, simple code completion. Think eager intern.
- Apple silicon 32GB+ (that Mac mini again): mid-size models run surprisingly well. This is the sleeper reason home-hosting people love the mini — vending machine and free model server.
- Big GPU (24GB+ VRAM): large models, genuinely capable — but now you bought a GPU and you're paying the power bill.
The honest summary: local models are ~1–2 years behind frontier hosted ones. Fantastic for private/bulk/background work and as a free tier inside your own product. Not what you want architecting your business logic.
A sane starter budget
| Item | Monthly |
|---|---|
| VPS | ~$6–12 |
| Domain | ~$1 (annualized) |
| One agent subscription (Claude Pro / Cursor / ChatGPT) | ~$20 |
| API budget for experiments (OpenRouter, capped) | $5–10 |
| Total | ~$32–43/mo |
That's the real entry price of a full AI build-and-host stack — less than most people's streaming subscriptions. Scale spending after something works, not before.
Next module: the decisions nobody can make for you — the pre-launch checklist, my own comfort-threshold story, and why all of this matters whether or not you ever charge a bot a cent.
📬 Optional: get the course by email
The whole course is free right here, no email needed — that stays true. But if you want it in your inbox, plus updates when modules are added or prices/tools shift, drop your email.
Disclosure: some links in this course are affiliate links (including Amazon). They help fund the free course — your price never changes, and every product here is one we'd recommend anyway.