Token Counter

Estimate how many tokens a passage of text is likely to consume. This is a heuristic approximation, not a tokeniser: it does not download any model vocabulary, and it cannot give exact counts for a specific model version.

Text to analyse

Nothing leaves the browser. Counting is done locally with the heuristic described in the sidebar.

Count spacing tokens Apply conservative (lower) estimate

Estimates by model family

waiting

Characters

0 excluding spaces

Words

0 unique

UTF-8 bytes

0 bytes per character

Characters per token

—

mean value across the text

Detailed estimate

Cost preview

indicative

Enter a per-million-token price to sanity-check a workload. Prices change often; these are your own numbers, not a quote from us.

Input $/1M

Output $/1M

Calls

Input cost (one call)

$0.0000

using the mid estimate

Total input cost

$0.00

1,000 calls

How to use

Paste the text you plan to send. Counts update as you type.
Read the mid column as your working figure and the range as the uncertainty band.
Long words, code, JSON and non-Latin scripts tokenise less efficiently than prose — the estimator accounts for this.
Turn off spacing tokens to see how much of the cost comes from whitespace and punctuation.
Use conservative estimate when you want a pessimistic budget figure.

Method & honest limits

The estimator segments the text into word, number, punctuation and whitespace runs, then applies a characters-per-token divisor per category and per model family, adding a small allowance for multi-byte scripts.

Typical English prose lands near 4 characters per token; this tool reports where your text falls.
Real tokenisers are trained on data, so any rule-based estimate drifts on unusual input.
Chinese, Japanese and Korean text is much denser per character — expect the higher end of the range.
Structural tokens, system prompts, tool schemas and chat framing are not included.
For billing-critical numbers, use the tokeniser published by the model provider.