How Many Tokens Are in a Word? GPT, Claude, Gemini Compared

The short answer: about 1.3 tokens per word in English. That holds across GPT, Claude, and Gemini for normal prose. The longer answer is more interesting — and matters if you are estimating costs, sizing context windows, or trimming prompts.

The Quick Conversion

English text	Approximate tokens
1 word	1.3 tokens
10 words	13 tokens
100 words	130 tokens
500 words	650 tokens
1,000 words	1,300 tokens
1 page (~250 words)	325 tokens
5 pages (~1,250 words)	1,625 tokens
1 short book (~50,000 words)	65,000 tokens

Or going the other way:

Tokens	Approximate words	Equivalent
100	75	One paragraph
500	375	Half a page
1,000	750	One full page
4,000	3,000	One short blog post
8,000	6,000	One long article
32,000	24,000	One short report
128,000	96,000	One short novel
200,000	150,000	One full novel

Why It Is Not Exactly 1 Token Per Word

Tokenizers do not split text on spaces. They split on a learned vocabulary of common subword pieces. This means:

• Common short words = 1 token. "the", "and", "is", "of" are each 1 token.
• Common longer words = 1-2 tokens. "computer", "function", "analysis" are 1-2 tokens each.
• Rare words = 3+ tokens. "hippopotamus" is 3 tokens. "antidisestablishmentarianism" is 7 tokens.
• Punctuation = 1 token each. Commas, periods, question marks each count.
• Spaces are usually attached to the following word in modern tokenizers.
• Numbers and code often use more tokens than text because they break into individual characters more often.

Tokens Per Word by Content Type

Content	Words per token	Why
Plain English prose	0.75	Standard tokenizer training data
Technical writing	0.65	More long words, jargon
Code (Python, JS)	0.50	Symbols and identifiers split heavily
Spanish/French	0.55	Different tokenizer coverage
Chinese/Japanese	0.35	Each character is often 1 token
Math/equations	0.30	Symbols are individual tokens
URLs and emails	0.25	Characters split heavily
JSON output	0.45	Brackets, commas, quotes all count

If you're processing code or non-English text, your token count will be much higher than the English rule of thumb. For Chinese or Japanese, 1 word is often 2-3 tokens.

Tokenizer Differences Across Models

Each major LLM uses a slightly different tokenizer. The same 1,000-word English document tokenizes to different counts:

Model	Tokenizer	Approx tokens (1,000 English words)
GPT-4o, GPT-4.1	o200k_base	~1,250
GPT-3.5, older GPT-4	cl100k_base	~1,300
Claude (all)	Claude tokenizer	~1,290
Gemini	SentencePiece	~1,280
Llama 3, 4	BPE	~1,310
DeepSeek	BPE variant	~1,295

The variation is small — within 5% — but it can matter at scale. If you are budgeting for 1 million prompts, a 5% difference is 50,000 tokens, which adds up.

How to Count Tokens Without the Math

The fastest way is to use a free online token counter:

1. Paste your text into the Token Counter
2. See instant count — tokens, words, characters
3. See estimated API cost across major models
4. Adjust your text and watch the count update live

The counter runs entirely in your browser — your text never leaves your device, so it works for confidential prompts and proprietary content.

When the Difference Actually Matters

For most casual use, 1.3 tokens per word is fine. The difference matters when:

• You are near a context window limit. If your document is 100K words and the model has a 128K token limit, you need to know whether you're at 130K or 145K tokens — exact counts matter.
• You are budgeting at scale. A 5% miscount across 1M API calls is 50K tokens — sometimes hundreds of dollars.
• You are processing non-English content. The English ratio breaks down for other languages, and assuming 1.3x can be off by 2-3x.
• You are sending code or structured data. Code uses ~2x the tokens of equivalent prose.

For everything else, the rule of thumb is fine. 1.3 tokens per word, 0.75 words per token. Use the counter when you need precision, use the rule when you need a rough estimate.