Saying Hi to Your AI Costs You Twice, and Seven Cheap Ways to Stop Paying for It

Someone types "hi" to your AI feature. You pay for the question. You pay again for the answer. Every AI feature runs on API keys and tokens. You pay for what goes in and for what comes out. A long chat resends its whole history every turn. So a million tokens, about a dollar, arrives fast. This lesson walks through the cheap fixes, live on my own AI gateway, Vatan. Cache the exact answer, so the second "hi" is free. Then cache by meaning, so "hello" gets the same answer too. Try the cheaper model, and let evals decide if it is good enough. In the demo the expensive model passed 90%. The cheap one passed 86% at a tenth of the price. Then the calculator. A million questions a month. Prompt caching alone takes the bill from $792 to $304. Each lever after that takes a bit more off. Last, batch requests. Most providers charge about half if you can wait up to 24 hours for the answer. Recorded as a Maven lightning lesson. What you will learn: - Why every AI call is priced twice, input and output - Exact caching vs semantic caching - How evals tell you if a cheaper model is good enough - Compressing chat history with the cheapest model - Using a small classifier model to find the user's intent - Where each lever lands on a real monthly bill - When a batch request cuts the price in half