LLM token counter

Models read text, and bill for it, in tokens: whole words, word pieces, single characters or raw bytes. Each vendor trains its own tokenizer, so the same text comes out as a different number of tokens on different models. Text that isn't English usually takes more tokens.

For open-weight models we use the tokenizer file the vendor published on Hugging Face, pinned to one version. Your browser downloads it once, keeps it, and counts in a background worker, so the number is what that model's own tokenizer gives for the same text.

When a vendor hasn't published its tokenizer, or before the files are loaded, the count is UTF-8 bytes ÷ 4, rounded up. That works out to about four letters per token for English and about 0.75 tokens per Chinese or Japanese character (three bytes each). Use it as a rough guide.

Only the text you paste. A real request also carries chat formatting such as role markers, plus any tool definitions or images, so the input tokens on your bill are a little higher. The usage field in each API response shows the billed numbers.

The cost is what this text costs as input, at the price you pay on cheapaimodel. A struck-through amount is the same text at the vendor's list price.

cost = tokens ÷ 1,000,000 × input price per 1M tokens

Counting happens on your device. The only network requests are for the tokenizer files, and they don't contain your text.