The quote everyone remembers is the build cost. The number that decides whether an AI feature survives is the monthly one, and it tends to show up in month two, right when the trial credits run dry and real traffic arrives.
I budget AI features the way I'd budget a delivery van. Buying it happens once. Fuel, insurance, and the driver run forever. Here's where the fuel actually hides.
Where the money goes
Tokens, including the ones you forget. Everyone counts the user's question. Fewer people count the system prompt, the retrieved context, the few-shot examples, and the model's own output, all billed on every call. A feature carrying a 3,000-token context that looked tiny in testing can run 10x the estimate once every request drags that prompt along.
Retries and retrieval. A retry on failure doubles the cost of that call. A RAG feature also pays to embed every document, store the vectors, and run a similarity search per query. The model bill is one line on a longer receipt.
