How Many Words Is a Token? Real Counts From Real Text

AI models read and write text in tokens, not words, and API prices are quoted per million tokens. The usual rule of thumb is that one token is about three quarters of an English word, so 1,000 tokens is roughly 750 words.

That rule is fine for ordinary prose and badly wrong for other kinds of text. Spreadsheets, invoice numbers, code and many non-English languages use far more tokens per word. If your work involves those, a cost estimate based on 0.75 words per token can be off by a factor of several.

Rather than repeat the rule, we ran five pieces of everyday office text through a real tokenizer and counted. The results are below, followed by what they mean for a monthly API bill. Last verified: November 1, 2026.

To turn your own word counts into a cost estimate, multiply tokens by the provider’s per-million-token price, as the worked examples below show.

How we counted

We used js-tiktoken, the JavaScript port of OpenAI’s tiktoken library, with two encodings:

• o200k_base: the encoding tiktoken maps to GPT-4o-family and newer OpenAI models.
• cl100k_base: the older encoding used by GPT-4 and GPT-3.5-era models, included for comparison.

Claude and Gemini use their own tokenizers, which are not published in this library. Anthropic’s documentation states that the tokenizer used from Claude 4.7 onward produces about 30% more tokens for the same text than earlier Claude models. So treat every count below as exact for OpenAI’s o200k_base and as an estimate for other providers.

Words were counted by splitting on spaces. Characters include spaces, line breaks and punctuation. The five samples were written for this test:

1. Business email: a 55-word reply to a supplier about a quote and delivery date.
2. Product description: a 34-word listing for a handmade mug.
3. CSV snippet: a header and four invoice rows (date, invoice number, customer, amount, status).
4. Code snippet: a four-line JavaScript function.
5. Traditional Chinese: one sentence asking a supplier to confirm a delivery date (我們下週二會寄出修改後的報價單，請確認交貨日期是否能在三月十四日之前。).

The results

*Each CSV row has no spaces, so it counts as one “word”. The useful figure is characters per token: 228 ÷ 101 = 2.3.

What stands out:

• Plain English prose lands close to the rule. The email came in at 55 ÷ 64 = 0.86 words per token and the product description at 34 ÷ 45 = 0.76. The common rule of 0.75 is slightly cautious for simple business writing, which is the safe direction for budgeting.
• Data is expensive. The email averages 332 ÷ 64 = 5.2 characters per token. The CSV averages 2.3, so the same amount of text costs more than twice as many tokens. Dates, amounts and IDs split into small pieces.
• Code sits in between. 24 ÷ 53 = 0.45 words per token, or 182 ÷ 53 = 3.4 characters per token.
• The newer tokenizer matters most for Chinese. The sentence dropped from 43 tokens on cl100k_base to 28 on o200k_base, 35% fewer. For the English samples the two encodings differed by at most one token.

Converting token limits into words

Using the measured ratios, here is what common token amounts hold. Each figure is tokens × words per token; pages assume 500 words per page.

Going the other way, 1,000 words of email-style prose is 1,000 ÷ 0.86 = 1,163 tokens on o200k_base, and 1,000 ÷ 0.75 = 1,333 tokens by the rule of thumb. On Claude, adding about 30% to the measured figure gives roughly 1,163 × 1.3 = 1,512 tokens. That is an estimate, not a measured count.

What this means for cost

Take a small online shop that wants an AI model to write 2,000 product descriptions like our sample. Each request sends a 150-word brief and gets back a 34-word description.

Using the measured product description ratio of 0.76 words per token:

• Input: 150 ÷ 0.76 = 197 tokens (rounded)
• Output: 34 ÷ 0.76 = 45 tokens, which matches the count we measured

On OpenAI’s GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens (OpenAI API pricing):

• Input: 2,000 × 197 = 394,000 tokens × $2 ÷ 1,000,000 = $0.788
• Output: 2,000 × 45 = 90,000 tokens × $10 ÷ 1,000,000 = $0.90
• Total: $1.69 for all 2,000 descriptions

The same job on Claude Sonnet 5.5, which has the same $2 / $10 list price (Claude API pricing), would cost about the same if token counts matched. If Claude’s tokenizer produces around 30% more tokens for this text, the estimate becomes 512,200 input tokens ($1.02) and 117,000 output tokens ($1.17), about $2.20. Still small, but a 30% gap matters at scale.

Now consider a job that sends spreadsheet data instead. Summarising 1,000 invoice exports of 20,000 characters each at the CSV ratio of 2.3 characters per token is about 8,696 tokens per export, or 8.7 million input tokens in total. At $2 per million, that is $17.39 in input alone. Estimating the same text with the prose ratio of 5.2 characters per token would give 3.8 million tokens and $7.69, less than half the real figure.

Practical rules

• For ordinary English emails and documents, budget 0.75 words per token. It slightly overestimates tokens, which keeps estimates safe.
• For spreadsheets, logs and anything full of numbers, count characters and use about 2–2.5 characters per token.
• For code, use about 3–3.5 characters per token.
• For Claude, add about 30% to counts made with OpenAI’s tokenizer as a cautious estimate.
• For anything you will run at volume, test a sample with the provider’s own token counter before committing to a budget. The usage figures in each API response are what you will actually be billed on.

Run your own numbers with the formula above and the official pricing pages linked in the sources, and see ChatGPT Plus vs Claude Pro vs Google AI Pro if a flat subscription might suit you better.

Written by Editorial team · Last updated: 2026-11-01 · Facts last verified: 2026-11-01 · Editorial policy