The point that a count from one encoding can't be reused for another model is easy to miss when estimating budgets. One practical addition: for chat-style requests, per-message overhead and tool or function schemas can add a surprising number of input tokens on every turn, so a words-to-tokens estimate on the user text alone tends to under-count. Comparing a local tiktoken estimate against the usage field returned by the API on a few real requests is a cheap way to calibrate that gap. Do you know how much the gap grows on code-heavy prompts?