How many tokens are used for an English word? Comparison of the differences between ChatGPT, Claude, and Gemini

When many people calculate the cost of AI Token, the first intuitive question is often: How many tokens will an English word consume?

Let’s talk about the conclusion directly: there is no fixed formula of “1 English word = several tokens” that is common to all three platforms. But if it’s in English, the rough estimate the official can give is much clearer than in Chinese. OpenAI officially wrote that 1 token is approximately equal to 0.75 English words; Google Gemini officially wrote that 100 tokens is approximately equal to 60 to 80 English words. The official document I found on Anthropic does not focus on "how many tokens per English word", but provides count_tokens for you to calculate directly.

Look at the table first: How many tokens are used for an English word?

I have deliberately divided the table below into two types of information: one is the extent to which the official statement is made, and the other is the range you can use when estimating costs in the early stage. Some of the "available ranges" are officially directly convertible, and some are conservative estimates for budget planning. I will mark them clearly.

Officially confirmed statement

Converted into approximately a few tokens per English word

The range that can be captured in early planning||ChatGPT / OpenAI

1 token ≈ 0.75 English words; also written as 100 tokens ≈ 75 words.

About 1.33 token / 1 word

About 1.2~1.5 token / word

Claude / Anthropic

The official focus is to provide Count Tokens API, which can first count the tokens of messages, including tools, images, and documents; the official documents I found do not directly provide the formula of "how many tokens per English word".

About 1.2~1.5 token / word (conservative practical estimate, unofficial fixed value)

Gemini / Google

100 tokens ≈ 60–80 English words; also write 1 token ≈ 4 characters.

About 1.25~1.67 token / 1 word

About 1.25~1.7 token / word

Why did I give a range, not a single number?

Because token is not the number of words, but the model’s own segmentation unit. OpenAI officials have directly stated that a token can be as short as one character or as long as a whole word; even the same word in English may be split differently in different contexts.

So even if they are all English words, the results you see may be affected by the following things:

Like a, the, apple, tokenization, they will not look the same. OpenAI officials even gave examples such as tokenization will be split into token and ization. This means that not every English word will correspond to exactly 1 token.

Are there any spaces, punctuation, or format before and after?||OpenAI officially states that spaces, punctuation, and some words will be counted into the token. In other words, word, word, word, and "word" are not necessarily the same number of tokens.

What is actually calculated is not just the word itself

If you send a request in the API, the things that will actually be included in the token calculation usually include system prompt, messages structure, tools, files, pictures, and historical conversations. OpenAI and Anthropic officially include these in the scope of token counting; Gemini also provides count tokens to count the entire input, not just the text.

ChatGPT: How many tokens does one English word have?

Of these three, OpenAI/ChatGPT has the clearest rough estimate in English. The official directly gives you two very useful formulas:

1 token ≈ 0.75 words

100 tokens ≈ 75 words

Putting this back, it is:

1 English word ≈ 1.33 tokens

This is very close to the number that you can use directly when making a budget.

If you want to be more practical, I would suggest:

ChatGPT early capture method

Ideal rough estimate: 1.2 token/word

Commonly used estimate: 1.3~1.4 token/word

Conservative budget: 1.5 token/word

常用估法：1.3～1.4 token / word

保守預算：1.5 token / word

The advantage of this is that you will not be too optimistic about the cost just because you only remember "0.75 word / token".

Claude: How many tokens does one English word have?

The biggest difference between Claude / Anthropic and OpenAI is: the official document I found does not focus on providing the formula of "how many tokens per English word", but tells you:

Use the Token Count API first

to directly count the tokens of the entire message

tools, images, and documents can also be calculated together

This means that Anthropic's official idea is more like: don't memorize the formula for each word, it is more reliable to count directly.

Claude’s most accurate statement here should be:

The key points that Claude can officially confirm

There is an official count tokens method

You can count the entire request first

But the official documents I found do not directly give a fixed formula of "1 English word ≈ several tokens"

Claude's early grasping method

If you are just doing a preliminary budget, I would suggest you grab it first:

This is a conservative estimate for budgeting, not Anthropic Official Guaranteed Value. Before it is officially launched, you should still run count tokens directly.

Gemini: How many tokens does one English word have?

The official statement from Gemini / Google is also very clear and very practical for English. Google officially writes directly:

1 token ≈ 4 characters

100 tokens ≈ 60–80 English words

If you push this sentence back directly, it is:

1 English word ≈ 1.25~1.67 tokens

So if you are using Gemini, it is very suitable to catch it like this in the early stage:

Gemini can be caught in the early stage

Ideal rough estimate: 1.25 token / word

Common estimation method: 1.3~1.5 token / word

Conservative budget: 1.6~1.7 token / word

This method is more in line with the actual usage situation than just remembering 4 characters = 1 token. Because when a request is actually sent, it usually does not contain only bare words.

If you just want to quickly capture the cost, the most practical table can look like this

This table is closer to the version you will use when actually making a budget:

In practice, it is safer to capture the number first

1 word ≈ 1.33 tokens

1.3～1.5 token / word

Don't memorize it by heart, count first

1.2～1.5 token / word

1 word ≈ 1.25～1.67 tokens | token

What really affects "how many tokens per English word" is not just the word itself

The most important thing about this article is not the formula, but this thing: the token you actually pay for in the end is usually not just the English word itself.

System prompt will also count

OpenAI's token counting and Anthropic's token counting count the entire message/request, not just the text.

If you retain historical context, previous messages may also be included in the calculation. Claude's context window document also specifically reminds you to estimate token usage first to avoid exceeding the context window.

Tools, pictures, and files will also be counted

Anthropic official clearly states that token count can include tools, images, and documents; Gemini official also has multi-modal tokens indicating the direction.

So if you just ask "How many tokens is an English word?", the answer can only help you make a very rough text budget at best. Once you enter the actual use of the API, what you really should look at is the entire request.

Conclusion: The 3 most memorable sentences from this article

First, it is easier to make rough estimates in English than in Chinese

Because OpenAI and Gemini officials directly give approximate values in English, it is indeed easier to estimate tokens in English than in Chinese.

Anthropic 官方明講 token count 可包含 tools、images、documents；Gemini 官方也有多模態 token 說明方向。

所以你如果只是問「一個英文單字多少 token」，那個答案最多只能幫你做很粗的文字預算。一旦進入 API 真實使用，真正該看的還是整份 request。

結論：這篇最值得記住的 3 句話

第一，英文比中文更容易做粗估

因為 OpenAI 和 Gemini 官方都直接給了英文近似值，英文 token 預估確實比中文更容易。

Second, both ChatGPT and Gemini can directly calculate the approximate range

ChatGPT: 1 word ≈ 1.33 tokens

Gemini: 1 word ≈ 1.25~1.67 tokens

Third, Claude’s best way is to directly count

Anthropic officials currently prefer to provide Count Tokens API instead of asking you to memorize every word formula.

So if I want to compress the whole article into one sentence, I would write it like this:

For an English word in ChatGPT, Claude, or Gemini, you can usually capture about 1.2 to 1.7 tokens first, but when it comes to officially calculating the cost, the official token counting of each platform shall prevail.

Is one English word necessarily equal to 1 token?

Not necessarily. OpenAI officials have made it clear that a token can be a single character, a partial word or a complete word, so English words are not fixed to one word and one token.

Which company is best for using English words to evaluate tokens?

ChatGPT and Gemini are relatively easy to estimate, because the official directly provides English rough rules. Claude recommends using count tokens directly.

Why does ChatGPT often say that 1 token is about 0.75 words in English?

This is a rough estimate rule in English provided directly by OpenAI, that is, 100 tokens is about 75 words.

Why does Gemini become about 1.25~1.67 tokens per word?

Because Google officially writes 100 tokens ≈ 60–80 English words, the inference is that each English word roughly falls in this token range.

Why didn’t Claude directly write how many tokens each English word has?

Because the official Anthropic document I found focuses on the Count Tokens API, which allows you to directly count the entire message first instead of providing a fixed words-to-token formula.

Data source and credibility statement

This article mainly refers to the official Token description of OpenAI, the official Count Tokens file of Claude and the official Token description of Gemini as the main source of information on the topic of "Approximately how many Tokens are used for an English word". Because the official documents of the three platforms are not written exactly the same way. For example, OpenAI and Gemini directly provide rough estimation rules in English, while Anthropic prefers to provide count tokens tools. Therefore, this article will write the official content and the range that can be used to estimate costs in practice to avoid mistaking the estimated values for fixed rules.

If you want to understand the differences between models, platforms and costs faster, you can also go back to AI Token to see the complete summary. The main purpose of this article is not to give you a single number that looks accurate but is actually unreliable, but to help you establish a judgment method that can be used to make estimates and know when to switch to the official count tokens tool.

This article belongs to the category "AI Token Computing".

This category mainly organizes how AI Token is calculated, the difference between input and output, the token consumption logic of different models or platforms, misunderstandings about word count and token conversion, backend usage interpretation and cost control concepts. It helps users who are new to AI API not only know that tokens will affect the price, but also better understand why the same English content may have different token performances on different platforms.

How many tokens are used for a Chinese character? Difference comparison between ChatGPT, Claude and Gemini

How many tokens will be consumed for the same content between ChatGPT, Claude and Gemini? Comparison of the differences between the three major platforms

How to view GPT Token billing? It is enough for novices to understand the key points first

API Token
Gemini Token
Claude Token
Token conversion
ChatGPT Token
English Token

AI Token organizes the basic concepts, calculation methods, API fees and model comparisons of AI Token (word elements), and covers common models such as ChatGPT, Gemini, Claude, etc. to help you establish clear understanding and judgment faster.

How many tokens are used for an English word? Comparison of the differences between ChatGPT, Claude, and Gemini