The LLM Cost Engineering Handbook

Most teams running AI in production can say what they spent last month. Far fewer can say what one resolved ticket or one generated report cost, or where the money went.

Take a typical agent job apart and about three-quarters of the bill buys no new information at all. It's the same text, sent again and again, because that's how a conversation with a model works. The model's price barely matters. Which means the obvious move, switching to a cheaper model, shouldn't be your first: it cuts the bill and it also changes what your product can do. Caching often saves as much or more and costs nothing in quality — you just have to find where it applies and prove it's working.

Sixteen chapters. Measure what you're paying for, apply six levers in the order they pay back, then keep the savings from creeping away. Five appendices: a fill-in cost model, a thirty-day plan, a forty-eight-point audit, a method for rebuilding every price when providers change them, and the source and date behind every provider figure in the book.

Every worked example states its assumptions, shows its formula and reconciles arithmetically — a script checks every figure against the text, and every statement about a provider carries a source and the date it was last checked. The examples are labelled as models, not dressed up as case studies. This is not an introduction to AI. It assumes you ship.

Two packages. The Book is the PDF and EPUB. The Book + Toolkit adds the AI Cost Model Toolkit: the fill-in worksheet, a calculator that runs every figure in the book against your own numbers, drop-in cost logging for Python and TypeScript, and the audit as a scored sheet.