Question 1

What is a token?

Accepted Answer

A token is the unit a language model actually reads — usually a common word, a word fragment, or a few characters. English prose averages roughly four characters per token, so 100 tokens is about 60–80 words. Models are billed and limited by tokens, not by words or characters, which is why counting them matters before you send a long prompt.

Question 2

Are these counts exact?

Accepted Answer

For OpenAI models, yes — exactly. OpenAI publishes its tokenizer tables, so we run the real thing in your browser and the number matches what their API would charge. Claude counts come from Anthropic's own official token-counting endpoint, so they are authoritative too. Gemini is the exception: Google publishes no tokenizer we can run, so that card is a clearly-labelled estimate.

Question 3

Why is the Claude number higher than OpenAI for short text?

Accepted Answer

Anthropic's endpoint counts a complete message, not a bare string, so it includes about nine tokens of role and structure framing that a real API request always carries. On one sentence that overhead is a large share of the total; on a long document it becomes negligible and the two counts converge to within a percent or two.

Question 4

How accurate is the Gemini estimate?

Accepted Answer

On English prose, usually within about 10%. We derive it from the exact OpenAI count rather than the common characters-divided-by-four rule, because Gemini and OpenAI both use large-vocabulary tokenizers and behave similarly on the same text — while a character ratio breaks down completely on code and non-Latin scripts. It remains an estimate, and the tool labels it as one.

Question 5

Why do non-Latin scripts count so differently?

Accepted Answer

Tokenizer vocabularies are built mostly from English text, so other scripts get split into far smaller pieces. The same Devanagari sentence can be 32 tokens for OpenAI and 79 for Claude — a 2.5x gap where English prose would differ by a few percent. The tool warns you when your text is mostly non-Latin, because an estimate derived from one tokenizer cannot be trusted to predict another there.

Question 6

Is my text sent to a server or stored?

Accepted Answer

OpenAI and Gemini numbers are computed entirely in your browser — that text is never transmitted. The Claude count does require a request to Anthropic's counting endpoint via our server, because no offline Claude tokenizer exists. Nothing is logged or stored either way, and if you would rather send nothing at all, the OpenAI count alone is fully local.

Question 7

Is it really free? Is there a limit?

Accepted Answer

Genuinely free, no account, and no per-use limit. Nothing here calls a paid API: OpenAI's tokenizer is open source, and Anthropic's token-counting endpoint is documented as free — it consumes no token quota. There is a generous rate limit purely to stop automated abuse.

Count tokens for ChatGPT, Claude and Gemini.

How it works

Frequently asked questions