Quick answer
A token is a small chunk of text, such as a character, part of a word, a whole word or punctuation, that a language model reads and writes. AI developers are billed per million tokens. This page covers language-model tokens, not crypto tokens. OpenAI's rough estimate for English is 100 tokens to about 75 words.
Key takeaways
- A token is a chunk of text a language model reads and writes. It can be a character, part of a word, a whole word or punctuation, per OpenAI's help center.
- Token counts are not word counts. OpenAI says the same text can produce different counts depending on the model, its encoding and the language.
- Tokens are the billing unit for developers who call a model through an API. Both vendors' pricing pages list prices per million tokens, with input and output priced separately.
- To estimate a cost: count your document's tokens with a vendor tokenizer, multiply by the vendor's per-million rate, and add the reply. The worked example on this page is our arithmetic.
- These are API prices. This page does not say what an app plan costs; check the plan's own page.
What is a token in AI?
A token is a small chunk of text, such as a character, part of a word, a whole word or punctuation, that a language model reads and writes. AI services count and bill usage in tokens. This page is about language-model tokens, not crypto tokens.
Think of tokens as the pieces a model cuts your text into before it works with it. OpenAI's help center says "A token can represent a character, part of a word, a whole word, or punctuation." Anthropic's glossary words it differently: "Tokens are the smallest individual units of a language model, and can correspond to words, subwords, characters, or even bytes (in the case of Unicode)."
Both vendors' definitions describe pieces that can be smaller than a word. OpenAI says "The same text can produce different token counts depending on the model, its encoding, and the language." Anthropic's glossary says the number of characters per token "can vary depending on the language used." So a token count is a measure of size for one model, not a universal unit.
How many words is a token?
For English text, OpenAI's rule of thumb is that 100 tokens are approximately 75 words, and 1 token is approximately 4 characters. OpenAI's help center calls these "estimates, not exact counts" and says other languages can have different relationships; its Tokenizer page says one token "generally" corresponds to about 4 characters for common English text.
Anthropic's pages give two rough figures. Its glossary says "For Claude, a token approximately represents 3.5 English characters, though the exact number can vary depending on the language used." Its pricing FAQ says "As a rough estimate, 1 token is approximately 4 characters or 0.75 words in English." That FAQ does not name a model.
Each is a rough vendor estimate, not a general fact. If you want a number for your own document, count it, as the steps below show.
| Vendor | Rule of thumb for English, as the vendor states it | Where |
|---|---|---|
| OpenAI | 1 token is approximately 4 characters; 100 tokens are approximately 75 words | Help center and Tokenizer page |
| Anthropic | For Claude, a token approximately represents 3.5 English characters (glossary); 1 token is approximately 4 characters or 0.75 words in English, as a rough estimate (pricing FAQ) | Glossary and pricing page |
How do tokens affect what I pay?
If you or your software calls a model directly through an API, tokens are the billing unit: the vendor charges a price per million tokens, and input and output are priced separately. OpenAI's help center says "Input, cached input, and output tokens can have different prices."
Two details can surprise you. OpenAI says "Reasoning tokens are not visible as answer text, but they count toward output usage and are billed as output tokens," so a short visible answer can use more tokens than it appears to. And OpenAI warns that "A lower price per million tokens does not necessarily produce a lower total cost," because models can tokenize the same text differently and generate different amounts of output or reasoning.
Anthropic's pricing page makes a different point about its own models: Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text." Its token counting page adds: "don't reuse token counts measured on the older model to estimate costs or context window fit." Compare the cost of your own task, not the price list alone.
These are developer prices. They are not what an owner pays for an app subscription, and this page does not state any app plan price.
What do the price lists say right now?
As of October 7, 2026, both vendors publish prices per million tokens in dollars, split into input and output. The table shows one model row from each vendor. We do not say the two models are equivalent or comparable in quality, and the table is not a recommendation; each vendor lists other models at other rates.
| Vendor and model (as of Oct 7, 2026) | Input, per million tokens | Output, per million tokens |
|---|---|---|
| OpenAI, gpt-6.1-sol (Standard tab, "Short context" columns) | $2.00 | $10.00 |
| Anthropic, Claude Sonnet 5.5 (model pricing table, "Base tokens" input and output) | $2 / MTok | $10 / MTok |
Both rows are standard rates: not batch, not cached input and not OpenAI's long-context columns. OpenAI's page, under "Prices per 1M tokens", has tabs for Standard, Batch, Flex and others; we read the Standard tab. Anthropic's page states "All prices are in USD." OpenAI's page shows dollar signs, and the part we read does not name a currency.
How do I estimate the cost of my own document?
Count the document's tokens with the vendor's tokenizer, multiply by the vendor's input price per million tokens, then add the expected reply at the output price.
- Get a token count for a representative page. OpenAI's help center says "Use the Tokenizer to see how text is divided into tokens," and its Tokenizer page is a tool where you paste text. Anthropic's token counting page says "Token counting lets you determine the number of tokens in a message before you send it to Claude." That is an API endpoint for developers, so ask whoever builds your tool to run it. The page calls the count "an estimate" and says "Token counting is free to use but subject to requests per minute rate limits based on your usage tier."
- Divide the token count by one million and multiply by the input price.
- Guess the length of the reply in tokens, divide by one million and multiply by the output price.
- Add the two, then multiply by how many documents you will run.
Here is a worked example. It is our arithmetic and an illustration built on assumptions we chose, not a vendor figure and not a typical cost.
- Assume a 1,500-word document. Using OpenAI's rule of thumb that 100 tokens are approximately 75 words, our arithmetic gives about 2,000 tokens (1,500 divided by 0.75). A real count will differ, because OpenAI calls the rule an estimate for English.
- At $2.00 per million input tokens (the gpt-6.1-sol row above), our arithmetic puts the input at $0.004.
- Assume a reply of 500 tokens, our own assumption. At $10.00 per million output tokens (the same row), that is $0.005.
- The total is $0.009 for one document, and $9 for 1,000 of them, on the price lists as of October 7, 2026.
Anthropic's glossary figure is in characters, so a conversion with it would need a character count and would give a different token total than OpenAI's rule. That is one reason to count instead of using a rule. The example also leaves out cached input, reasoning tokens and long-context rates, which the vendors price differently.
Do I need to know this?
Probably only if you or someone you hire builds a tool that calls a model through an API. These are API prices, and the pages we read do not say what an app plan charges or how an app passes costs on. For an app, this site suggests three questions for the seller: is the price per use or flat, does it cover the replies as well as what you send, and is there a cap.
Tokens matter to an owner when a per-use bill is quoted and you want to sanity-check it. A per-use charge that depends on how long your documents and replies are is the same idea, in a different unit, as the per-task billing covered in why Zapier feels expensive. For a worked per-call bill in another AI product, see AI receptionist cost. For where AI tools fit a small business, start with the practical guide to AI automation.
FAQ
How many words are 1,000 tokens?
OpenAI's help center gives the rule of thumb that 100 tokens are approximately 75 words for English text, and calls these estimates, not exact counts. Our arithmetic on that figure is about 750 words for 1,000 tokens. Anthropic's glossary gives about 3.5 English characters per token for Claude, and its pricing FAQ gives a rough estimate of 4 characters or 0.75 words per token in English.
Do AI tokens cost money?
For developers who use a model through an API, yes: OpenAI and Anthropic both publish prices per million tokens. For a chat app you subscribe to, this page makes no claim about what your plan costs or how it limits use. Read the plan's own page for that.
How much does a million tokens cost in ChatGPT?
Prices per million tokens are API prices and depend on the model. As of October 7, 2026, OpenAI's pricing page lists one model, gpt-6.1-sol, at $2.00 per million input tokens and $10.00 per million output tokens on its Standard tab, short-context columns, and others at different rates. That is a developer price, not a ChatGPT subscription price.
Is an AI token the same as a crypto token?
No. This page covers language-model tokens, the chunks of text an AI model processes. Crypto tokens are a separate subject that this page does not cover.
Do two AI models count the same text the same way?
Not necessarily. OpenAI says the same text can produce different token counts depending on the model, its encoding and the language. Anthropic says its Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text.
Who pays for the tokens when I use an AI app?
The pages we read describe what developers pay when they call a model directly. How a given app or plan passes those costs on is not stated on the pages we read, so this page makes no claim about it. Ask the app's seller what is included in your price.
What this page covers and does not cover
This page covers language-model tokens: the chunks of text that AI models read and write and that developers are billed for. It does not cover crypto tokens, how a model uses tokens internally or how tokenizers are built.
Prices and model names change often, and the figures here are tied to the date they were read. Open the linked pricing pages before you decide anything.
Drafted with Claude, a model made by Anthropic, one of the two vendors covered here; checked against each vendor's published material, as read on October 7, 2026. Method: definitions and rules of thumb come from OpenAI's help center and Tokenizer page and Anthropic's glossary and token counting page; prices come from each vendor's pricing page. OpenAI's help article was read rendered in a headless browser because a plain fetch is blocked. Each vendor's statements are reported as its own and were not independently tested here.