What counts as a token on your bill

Models do not read letters or words. They read tokens: common words whole, rarer words in pieces, punctuation and spaces. As a rough guide for English, a token is a little shorter than a word. Code, numbers and other languages split differently.

Each model family has its own way of splitting text. The same paragraph is a different number of tokens on different models. So a lower price per token is not always a lower price per page.

Your message is a small part of the input. The system prompt is sent with every request. So is the whole conversation so far. So are the definitions of any tools the model may call, and the results those tools return.

On the output side, a model that reasons before answering may write far more than it shows. That reasoning is commonly billed as output.

Models that accept images or video turn them into tokens too. How many depends on the model and on the size of the image. A few pictures can outweigh pages of text. If you send images, read how the host counts them.

Do not estimate from words. Send real requests and read the usage figures in the response. They give input tokens and output tokens, and often cached tokens, for that request.

Log them for a day. The totals are the numbers to multiply by the prices on this site.

Shorten the system prompt. Trim old turns from long conversations, or summarize them. Send the part of a document that matters, not the whole file. Cap the output length. Each of these lowers the count, whatever the price.

Runcrate serves 125 models through one OpenAI-compatible API. Every response reports its token counts, and one prepaid balance shows what they cost.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.