What Is a Token in AI? Why It Decides What You Pay
A token is the unit a language model reads and writes: a word, part of a word, or punctuation. About 100 tokens equal 75 English words. Full guide.
By Galactic News Editorial · Published 2026-11-10T01:25:54+00:00 · Updated 2026-11-10T01:25:54+00:00
What Is a Token in AI? Why It Decides What You Pay
A token is the unit a language model uses to process text. It can be a whole word, part of a word, a single character, or a punctuation mark. Models do not read words the way you do. They read tokens. Every prompt you send and every answer you receive is measured in tokens. That count sets how much fits in a conversation, and it sets what you pay when you use an AI service through an API.
What Is a Token in AI, Exactly?
Before a model sees your text, a program called a tokenizer cuts it into pieces. Each piece becomes a number. The model works only with those numbers. It then writes its answer as new numbers, which turn back into text you can read.
Common words often become one token. Long or rare words split into several. The word "tokenization" might split into "token" and "ization". Spaces matter too. A word with a leading space can tokenize differently from the same word without one. Capitalization changes the split as well.
There is no single token count for a piece of text. Each model family trains its own tokenizer. The same paragraph produces different counts on different models. The count also changes with language. English is efficient. Languages with less training data can cost many more tokens for the same meaning.
How Many Words Is a Token?
OpenAI gives three rules of thumb for English text. One token is about 4 characters. One token is about three-quarters of a word. One hundred tokens are about 75 words. These are estimates, not exact counts.
Use this table as a rough guide for English prose.
| Tokens | Approximate words | Approximate characters |
|---|---|---|
| 100 | 75 | 400 |
| 1,000 | 750 | 4,000 |
| 10,000 | 7,500 | 40,000 |
| 100,000 | 75,000 | 400,000 |
Code, JSON and formatted text tokenize less efficiently than plain prose. A short prompt with heavy formatting can use more tokens than a longer plain paragraph. Always count before you assume.
The Four Kinds of Tokens on Your Bill
Token pricing is not one number. Providers split tokens into categories, and each category can cost a different amount.
- Input tokens. Everything you send: your question, pasted documents, system instructions and chat history. Also called prompt tokens.
- Output tokens. Everything the model writes back. Also called completion tokens. Output tokens usually cost more than input tokens.
- Cached input tokens. Input tokens reused through prompt caching. They often cost less than fresh input tokens.
- Reasoning tokens. Tokens a reasoning model uses internally before producing its visible answer. They are not shown as text, but they count toward output usage and are billed as output tokens. A short visible answer can use far more tokens than its text suggests.
Why Tokens Decide What You Pay
API providers price models per million tokens, with separate rates for input and output. OpenAI's pricing page lists GPT-4o at $2.50 per million input tokens and $10.00 per million output tokens on standard processing. The formula is simple: tokens used times price, divided by one million.
Take a 2,000-word prompt with a 1,000-word reply. At roughly 0.75 words per token, that is about 2,667 input tokens and 1,333 output tokens. On GPT-4o rates, the prompt costs about $0.0067 and the reply about $0.0133. Small numbers per request, but they add up across thousands of requests.
A lower price per million tokens does not always mean a lower total cost. Models tokenize the same text differently. One model may generate more output or more reasoning tokens to finish the same job. Compare total cost on representative tasks, not just the headline rate.
Tokens and the Context Window
Every model has a context window. It is the maximum number of tokens a request can hold, including your prompt and the model's reply. The newest models from the major providers take around a million tokens. Older or smaller models take far less.
When a prompt is too long, you have options. Shorten or rephrase it. Remove repeated context. Split the work into smaller parts. Summarize documents before sending them. Leave room for output tokens too, especially with reasoning models, which need space for their internal thinking.
How to Count Tokens Before You Send a Request
Do not guess. Count.
- Use OpenAI's Tokenizer tool to see exactly how a piece of text splits into tokens.
- For code, use the tiktoken library. Pick the encoding for your target model, for example with tiktoken.encoding_for_model(model).
- Check the usage fields in the API response after each call. Chat Completions reports prompt_tokens, completion_tokens and total_tokens.
- Review totals over time in the provider's usage dashboard.
Remember that a plain-text count does not always match the full request count. Message structure, tools, schemas, images and files can add tokens. Count the complete request when cost matters.
How to Keep Token Costs Down
A few habits cut token spend without hurting quality. Send only the context the model needs. Trim chat history to the last few turns. Use cached prompts for repeated instructions. Ask for short answers when a short answer will do. Pick the smallest model that handles the task well. Test on a sample before you run a large batch.
Bottom Line
A token is the working unit of a language model: a word, a word fragment, a character or a punctuation mark. Roughly 100 tokens equal 75 English words. Tokens set the context limit and drive the bill, with input, output, cached and reasoning tokens priced separately. Count tokens before large jobs, and you will never be surprised by the cost.