Every API call to a language model is billed by the token, not the word, and models split text in ways that surprise you (unbelievable can be three tokens; a stray space is one more). This tool counts them exactly, using each model’s real tokenizer, and estimates the bill.
Paste anything, a system prompt, an article, a whole conversation. Nothing leaves your browser.
How to read it
- Tokens is the exact count from the model’s byte-pair tokenizer. GPT-4o and o1 use
o200k_base; GPT-4 and 3.5 usecl100k_base. - Cost / call is your input tokens at the model’s input price, plus the assumed reply length at its output price. Change the reply length to match what you actually expect back.
- Open Show the tokens to see the exact pieces the model reads, the same view behind the atoms of a model.
Prices are approximate and change often, so treat the numbers as a close estimate rather than a quote. Claude and Gemini use their own tokenizers, so their real counts differ a little from the OpenAI numbers shown here.