Blog/September 15, 2026·5 min read

How Many Tokens Is 1,000 Words? (About 1,350 Tokens)

Quick Answer

1,000 words of English prose is about 1,350 tokens with OpenAI's o200k_base tokenizer. Claude uses more tokens for the same text, roughly 1,550 to 1,820. Text full of numbers, code, or a language other than English needs noticeably more.

The popular rule says 1,000 tokens is about 750 words. We wanted a measured number, so we ran OpenAI's o200k_base tokenizer over our own published articles. Plain English prose came to 1,348 tokens per 1,000 words, which matches the rule closely. Articles packed with figures, like our page-count guides, came to 1,483.

Tokens for Common Word Counts

WordsGPT tokens (o200k)Claude tokens (estimate)
100135155 to 182
250337388 to 455
500674775 to 910
7501,0111,163 to 1,365
1,0001,3481,550 to 1,820
1,5002,0222,325 to 2,730
2,0002,6963,100 to 3,640
5,0006,7407,751 to 9,099
10,00013,48015,502 to 18,198

Claude figures are a range because Anthropic does not publish Claude's tokenizer. OpenAI-style tokenizers undercount Claude, so the range runs from 1.15 to 1.35 times the GPT count. Use Anthropic's token counting API when you need the exact figure.

Language Changes the Count

Tokenizers are trained mostly on English, so other languages split into more pieces. We measured the same three articles in eight languages. These articles contain many numbers, so every row sits above the plain-prose figure, but the differences between languages are what matter:

LanguageTokens per 1,000 wordsvs. English
English1,508baseline
French1,547+3%
Spanish1,563+4%
Portuguese1,571+4%
Indonesian1,905+26%
Arabic2,000+33%
Turkish2,146+42%
Polish2,316+54%

Spanish, French, and Portuguese cost only a few percent more than English. Turkish and Polish, with long words built from many suffixes and endings, cost 40 to 55% more. If you pay per token or work close to a context limit in these languages, budget for it.

How We Measured

  • Tokenizer: OpenAI's o200k_base encoding, run with the open-source gpt-tokenizer package.
  • Text: the body of our live articles, with navigation, scripts, and tables removed. Prose figures come from seven essay-style articles totaling 8,418 words.
  • Words: counted by splitting on whitespace, the same way most word counters work.
  • Date: September 2026. Tokenizers change between model generations, so newer models may count differently.

Related Conversions

Going the other way? See how many words fit in 1,000 tokens. To check a real prompt, paste it into the Token Counter, and use the Word Counter for the word side.

Sources

  1. OpenAI. tiktoken: BPE tokeniser for use with OpenAI's models (o200k_base encoding).
  2. gpt-tokenizer: JavaScript implementation of OpenAI's tokenizers.
  3. Anthropic. Token counting documentation.

Count the exact GPT tokens in your own text and estimate the Claude count.

Open Token Counter

Related Guides

Frequently Asked Questions

About 1,350 tokens for English prose with OpenAI's o200k_base tokenizer, used by GPT-4o and later models. Claude typically needs more, roughly 1,550 to 1,820 tokens for the same text.

On average about 1.35 tokens per English word in prose. Short common words are one token, while long words, numbers, and punctuation push the average up.

Roughly, for English. In our measurement English prose averaged a little over 4 characters per token. Other languages average fewer characters per token, so the rule overestimates how much text fits.

Tokenizers split long numbers into short digit groups, and commas and decimal points become separate pieces. Our number-heavy articles averaged about 1,480 tokens per 1,000 words versus 1,350 for plain prose.

Paste it into a token counter. For OpenAI models the count can be exact in the browser. For Claude, use the token counting endpoint in Anthropic's API, because Claude's tokenizer is not public.