A complete cost-engineering guide: model routing, prompt caching and batch discounts, subscription-vs-API arbitrage, hidden inflations (tokenizer changes, repricing), and a worksheet for costing an AI feature before building it.
Every AI feature in your business has a per-unit cost, and most owners can't see it. This paper lays out the cost-engineering playbook the sharpest builders are using in 2026: route cheap models for execution and expensive ones only for planning (builder Amir's 44¢-vs-$2.38 math); use subscriptions where they beat metered rates (Theo Browne's finding that a $200 subscription buys roughly $8,000 of inference at API rates); budget for what a heavy day actually burns (writer Frédéric Auguste's $30–60 of list-price compute); and watch for hidden inflation like tokenizer changes (Simon Willison's flag of a ~40% hidden price increase on Opus 4.7). It closes with a worksheet that costs an AI feature before you build it. Every number below is sourced to a named builder or analyst. No invented figures.
Summary drawn from the paper's executive summary (Human-N-Loop research series, October 2026 draft).
lead magnet for the $97 audit; credibility piece for cost conversations.
Sources named in the research: Auguste, Amir, Willison, Theo Browne, Nate B. Jones.