How Many Tokens Is a Word? We Measured It

How many tokens is one word?

About 1.3 on OpenAI's current tokenizer, for ordinary English. We ran five samples through o200k_base, the encoding GPT-5, GPT-4.1, GPT-4o and o3 bill with, and through the older cl100k_base. English fiction came to 1.32 tokens per word, a technical article to 1.39. OpenAI's own rule of thumb, 0.75 words per token, works out to 1.33, so it holds for plain prose.

Measured 6 Nov 2026 with js-tiktoken 1.0.21. A word is anything between spaces, the same rule our word counter uses. How we picked the samples is at the end.

The technical article costs more per word than the novel because it is full of model names, prices and numbers such as "GPT-5" and "$1.25", and each of those splits into several tokens.

How many tokens is 1,000 words?

Between 1,300 and 1,400 tokens for English on OpenAI's tokenizer: 1,318 for our fiction sample and 1,390 for the article. The same 1,000 words in Spanish came to 1,424 tokens.

Turned around, 1 million tokens holds about 758,000 words of plain English prose, or about 720,000 words of technical writing. For a quick estimate of your own text, multiply its word count by 1.3 for prose and 1.4 for technical writing, or paste it into the OpenAI token counter for the exact number.

Do Claude and Gemini count words differently?

Yes. Gemini matches OpenAI on plain prose but uses 9% to 19% more tokens on technical writing, code and JSON. We sent the same five samples to Google's countTokens method for Gemini 3.8 Flash, which counts tokens without running the model:

Tokens per word, measured 6 Nov 2026. Google's own rule of thumb, about 4 characters per token, works for the novel but undercounts the rest: it predicts 1.46 tokens per word for our article, and Gemini counted 1.52.

Anthropic's own figures put Claude above both. It says a token on its earlier models is about 3.5 English characters, and that Claude 4.7 and later produce about 30% more tokens for the same text. We did not run our samples through Claude, so the rows below are estimates: those published ratios applied to our two English samples.

Why do code, JSON and other languages cost more tokens?

A tokenizer gives its own token to pieces of text that are common, and common English words are very common, so most of them are a single token. Anything rarer gets split into more pieces. In code, brackets, quotes, dots and indentation tend to become tokens of their own, and long identifiers break into parts. JSON is the extreme case: every key comes wrapped in quotes and followed by a colon, so our sample used 4.13 tokens per whitespace-separated word.

Language matters too, and so does the tokenizer generation. Our Spanish sample needed 1.42 tokens per word on o200k_base but 1.63 on the older cl100k_base, so moving from GPT-4 to a GPT-4o or GPT-5 model cut Spanish token counts by about 12% before any price change.

If a prompt carries a lot of JSON, it is worth trimming keys and whitespace before you send it; our guide to cutting OpenAI API costs covers that and prompt caching.

What does 1,000 words cost to send?

Multiply the token count by the input price per 1M tokens. At 1,390 tokens per 1,000 words of technical English, a model that charges $2.00 per 1M input tokens, such as GPT-6.1 Sol, bills $0.00278 for those 1,000 words. Replies cost more, because output tokens are priced several times higher than input on most models.

Every current model's input, output and cached price is on our LLM API pricing page, and the LLM cost calculator turns a token count and a request volume into a monthly bill.

How we measured

• English fiction: the first 2,000 words of Pride and Prejudice from chapter 1, from Project Gutenberg (public domain).
• English technical article: the full body of our OpenAI cost-cutting guide, 1,388 words.
• Spanish fiction: the first 2,000 words of Don Quijote from chapter 1, also from Project Gutenberg.
• TypeScript code: one source file from this site, src/data/pricing.ts.
• JSON: the first 12,000 characters of one of our pricing snapshots.

Words are counted by splitting on whitespace. OpenAI tokens come from js-tiktoken 1.0.21 with the o200k_base and cl100k_base encodings, which match OpenAI's own tiktoken. Gemini tokens come from the Gemini API's countTokens method for gemini-3.8-flash, which adds nothing of its own: a one-word message counts as 1 token. The script is in our repository as scripts/measure-tokens.mjs, so anyone can rerun it, and it adds Claude and Gemini counts from their free token-counting APIs when keys are set. The article and code samples are pages and files from this site, so a rerun on a later version gives slightly different counts for those two; ours come from the 6 Nov 2026 versions.

To check the numbers outside the script, we pasted the fiction sample into our token counter. It gave the same 2,637 tokens for GPT-6.1 Sol.