Business SystemsGuide
How Many Tokens Is a Word? The Number That Quietly Decides Your AI Bill
AI models are priced per token, but people write in words. The conversion is simple enough to do in your head, and it explains why the same prompt can cost very different amounts on different models.
Author
DueClix Engineering
Published
Reading time
3 min read
Every AI model you pay for by usage (Claude, GPT, Gemini) bills in tokens, not words. A token is a small piece of text: often a whole short word, sometimes part of a longer one, sometimes a space or a punctuation mark. You never see them, but they are what the invoice counts.
The rule of thumb
OpenAI's own guidance for English text is that one token is about four characters, or about three quarters of a word. Turned around:
| Text | Words | Tokens (approx.) |
|---|---|---|
| A short email | 100 | 130 |
| A one-page document | 500 | 670 |
| A long help article | 1,000 | 1,330 |
| A 20-page contract | 10,000 | 13,300 |
That is close enough for planning. It is not exact, and it drifts in three predictable ways.
Three things that change the count
- The model. Each vendor uses its own tokenizer. Anthropic's pricing page says its newer Claude models produce about 30% more tokens for the same text than its older ones, so the 1,000-word article above is closer to 1,730 tokens there.
- The language. Tokenizers are built mostly from English, so text in Urdu, Arabic, Hindi, Chinese and many other languages usually takes noticeably more tokens for the same meaning. How many more depends on the language and the model; measure it on your own text rather than assuming.
- The content. Code, tables, long numbers, URLs and unusual names break into more pieces than ordinary prose.
Why it matters: a worked example
Take a support assistant that answers 10,000 questions a month. Every request sends its instructions plus a few help articles: about 3,000 words, or roughly 4,000 tokens by the rule of thumb (about 5,200 on newer Claude models). Input only, standard list prices, no caching:
| Model | Input tokens a month | Input cost a month |
|---|---|---|
| GPT-6 Luna (small) | 40M | 40M × $0.10 = $4.00 |
| Claude Haiku 5.5 (small) | ~52M | 52M × $0.10 = $5.20 |
| GPT-6.1 Sol (mid-range) | 40M | 40M × $2 = $80 |
| Claude Sonnet 5.5 (mid-range) | ~52M | 52M × $2 = $104 |
Two lessons sit in that table. First, the tier of model matters far more than the vendor: the small models are twenty times cheaper per token than the mid-range ones. Second, the token count is a multiplier on everything. Halve the words you send with each question and every line in the table halves with it.
How to count your own
- For planning: paste a real prompt into our free token counter. It estimates the tokens and shows what that prompt costs to send on ten current models.
- For the monthly picture: the AI cost calculator adds answers, caching and batch pricing.
- For billing-grade numbers: use the vendor's own token counting on your actual requests, or read the token counts your system already logs.
The conversion from words to tokens is the least interesting part of AI pricing, and the one most worth getting right. Everything else on the price list gets multiplied by it.
Sources
- OpenAI, What are tokens and how to count them?. The four-characters and three-quarters-of-a-word rules of thumb for English.
- Anthropic, Pricing, accessed 12 October 2026. The note that newer Claude models produce approximately 30% more tokens, and Claude list prices.
- OpenAI, API pricing, accessed 12 October 2026. GPT-6 Luna and GPT-6.1 Sol list prices.