LLM cost measurement

What we found when we checked our own numbers against what actually happened. One audit compared every figure on Vigil’s dashboard with the code that computes it: nine definitions were wrong, including a savings total that was really a rolling window.

Another began with a single call that cost three cents more on the customer’s invoice than in our figures, and found two kinds of cost that token-based metering never sees: per-request fees such as web search, and tokens billed outside a response’s top-level usage. Each post says how the finding was checked, and what to look for in your own figures.