Token Cost Estimator From Pasted Text

The useful thing about pricing a specific piece of text is that it turns an abstract per-million rate into a number with a decimal point you can reason about. A rate of a few dollars per million tokens sounds like nothing until you notice you are sending a 40,000-token document on every one of a hundred thousand requests, at which point the same rate is a five-figure line on a monthly invoice.

The characters-per-token ratio is the whole accuracy story here, and it is deliberately a field rather than a constant. English prose sits near four characters per token for most modern tokenizers. Code sits lower, because punctuation and indentation fragment into single tokens. Languages written in non-Latin scripts sit lower still — often dramatically — because the tokenizer's vocabulary was built mostly from English text, so the same sentence costs more tokens, and therefore more money, in Japanese or Hindi than in English. If you serve a multilingual product, run this once per language before you assume one average.

What the estimate leaves out is everything wrapped around your text. The chat template adds role markers and separators. Tool definitions are serialised into the prompt and can be larger than the prompt itself. On a multi-turn conversation, the entire history is resent on every turn, so the cost of the tenth message includes the first nine. None of that appears in a character count of the box above, and all of it appears on the bill.