Ever noticed that AI price lists never mention words or pages? They count something smaller.
Tokens in Everyday Terms
Think of a token as a word-piece. Common short words are often a single token. Longer or rarer words get split into two or more pieces. Spaces, punctuation and capital letters change how text is split.
The providers give similar rules of thumb for English:
- OpenAI says one token is about 4 characters, or roughly three quarters of a word, so 100 tokens are about 75 words.
- Google says that for Gemini models a token is about 4 characters, and 100 tokens are about 60 to 80 English words.
So for planning purposes: 1,000 tokens is roughly 750 English words, about one and a half pages of plain text. Both companies stress these are estimates. The same sentence can come out as a different number of tokens on different models, because each model family has its own tokenizer: the rule book that cuts text into pieces.
Myth vs Fact
| Myth | Fact |
|---|---|
| One token is one word | In English a token is about three quarters of a word, on OpenAI’s and Google’s rules of thumb |
| Every model counts tokens the same way | Each model family has its own tokenizer, so counts differ a little |
| You only pay for the text you can see | OpenAI says hidden reasoning tokens still count as output |
Why Other Languages and Code Use More Tokens
Tokenizers are built from large amounts of text, and they learn the most common pieces. Text that looks unlike that training mix tends to split into more, smaller tokens. OpenAI’s help page notes that other languages can have a different relationship between characters, words and tokens.
In practice, that means:
- Non-English text often needs more tokens for the same meaning. Check with a tokenizer tool before you estimate costs for another language.
- Code, numbers and unusual formatting (long IDs, tables, symbols) can use more tokens than ordinary prose of the same length.
Why Tokens Matter to You
1. They Set the Limits
Every model has a context window: the most tokens it can handle in one request, counting your instructions, the conversation so far and its own answer. On our price board, 36 of the 42 models we track list a window of about 1 million tokens. Chat apps also limit messages and length in ways that come down to tokens.
2. They Set the Price
AI APIs charge per token, usually quoted per 1 million tokens. Text you send (input) and text the model writes (output) have separate prices, and output costs more. On our board on 2026-11-21:
| Measure (per 1M tokens) | Input | Output |
|---|---|---|
| Cheapest model | $0.10 | $0.20 |
| Median across 42 models | $1.75 | $8.75 |
The two cheapest figures come from different models: GPT-6 Luna and Claude Haiku 5.5 list the lowest input price, while Ministral 3 14B charges the same $0.20 for input and output. At the median prices, a request with 1,000 tokens in and 500 tokens out costs about $0.006, a little over half a cent. Our guide How LLM API Pricing Works explains the full picture, including discounts for cached and batch requests.
3. Hidden Tokens Count Too
OpenAI’s help page explains that reasoning models can use internal reasoning tokens that you never see as answer text, and these still count as output. So a short visible answer can cost more than its length suggests.
Did you know? At the median input price on our board ($1.75 per 1M tokens), sending this whole article to a model as input would cost well under one cent.
How to Count Tokens
You rarely need exact counts. When you do, use the provider’s own tools:
- OpenAI offers a free web tokenizer tool. Paste text and it shows how the text is split and how many tokens it uses.
- Anthropic offers a token-counting endpoint for Claude. Its documentation says it is free to use (with rate limits) and that the count is an estimate that can differ slightly from the final bill.
- Google documents ways to count tokens for Gemini in its developer guide.
- Every API response reports the tokens actually used. That is the most reliable number for budgeting.
A Quick Way to Estimate
- Count the words in a typical request (your instructions plus the user’s text plus any document).
- Divide by 0.75 to get approximate tokens.
- Do the same for a typical answer.
- Multiply by the input and output prices on our price board, then by your number of requests.
For example, a 600-word email plus 150 words of instructions is about 1,000 input tokens. A 300-word reply is about 400 output tokens.
How many words was the longest prompt you wrote this week, and roughly how many tokens would that be?
What This Means for You
- Developers: use the provider’s token-counting tools while you design prompts, and log the token counts each API response returns.
- Founders and team leads: budget in tokens, not words, and expect other languages and code to use more tokens than English prose.
- Curious beginners: a token is roughly three quarters of a word. You rarely need an exact count to use AI tools well.
Bottom Line
Tokens are the word-pieces that AI models process, and the unit that limits and prices everything they do. As a rule of thumb, 1,000 tokens is about 750 English words; other languages and code usually take more. To go further, our Start here if you’re new to AI collection lists beginner-friendly sources, and How large language models work inside goes deeper into the mechanics.
Next, see what tokens add up to in practice in What Does One AI API Request Cost?. Want these updates by email? Sign up here, or follow new guides and data stories through our RSS feed.
