Tokens per Word, Measured: English and 11 Other Languages

Jonhisking · October 2, 2026

"One token is about four characters, or three-quarters of a word." You will find this rule of thumb everywhere. It comes from OpenAI's older tokenizers, and for modern models it is a little pessimistic for everyday English and wildly wrong for many other languages. We measured real text on GPT's current tokenizer (o200k, used by GPT-4o and later) to give updated numbers.

English: about 1.1 to 1.3 tokens per word

We counted four kinds of English text:

So for everyday English on current models:

• 100 words ≈ 115 tokens (plain prose) to 133 tokens (technical writing)
• 1 token ≈ 4.4–4.9 characters
• 1,000 tokens ≈ 750–890 words

The classic "¾ of a word per token" (1.33 tokens per word) matches technical text, so it is a safe, slightly conservative estimate. For ordinary writing, you will usually use about 15% fewer tokens than it predicts.

Code is denser: in our measurements, Python and JavaScript averaged about 3.5 characters per token. See How many tokens does code use?

Other languages: tokens per character

Words are a poor unit across languages: Chinese, Japanese and Thai do not put spaces between words, and languages like Finnish or Turkish pack a whole phrase into one long word. Characters per token is more useful. Here is the same customer-support prompt in several languages:

Two things stand out:

• Characters per token tells you little about cost on its own. Chinese has only 1.4 characters per token, yet costs almost the same as English, because each Chinese character carries much more meaning. The last column, tokens for the same meaning, is what matters for your bill.
• For Korean and Japanese, plan on roughly one token per 1.2–1.8 characters, and expect about 1.4× (Korean) and 1.8× (Japanese) the tokens of the same content in English.

The full results for 41 languages are in our language comparison.

Quick conversion table (English, current GPT models)

Other models count differently

These numbers are for OpenAI's o200k tokenizer. Claude and Gemini use their own tokenizers, and the same text can be noticeably more tokens on them. Anthropic's newer Claude models in particular use more tokens for the same text than GPT. See How to count tokens for GPT, Claude and Gemini.

Count your own text

Rules of thumb are fine for rough planning. For anything you will pay for, paste your real text into the token counter: it runs the actual tokenizer in your browser and shows the cost on every model.