What will this actually cost?

Getting the numbers right

Input and output tokens

Tokens are not words. A rough English estimate is ~4 characters per token, but that breaks down badly on code and on non-English text. If the number matters, count it properly — every major provider ships a token-counting endpoint, and using one is far more reliable than a rule of thumb or a third-party tokenizer built for a different model family.

Cached input

Caching is the largest lever on this page, and the most commonly left on the table. If every request begins with the same long system prompt, tool definitions, or retrieved document, that prefix can usually be cached and re-read at a fraction of the input price. Move the slider and watch what happens to the total — on a long-prefix workload it is often the difference between two models being close and one being obviously right.

The catch is that caching is a prefix match. One changed byte early in the prompt — a timestamp, a re-ordered JSON key, a session ID — invalidates everything after it. A cache hit rate you assumed but never measured is worth checking before you budget against it.

What this does not include

Batch discounts, committed-use agreements, enterprise pricing, fine-tuning, and per-request minimums are all out of scope here. Treat the output as the list-price ceiling for your workload, not a quote.