Tokens Are Not Characters: How to Read AI Pricing and Input Limits

A token is the “chunk of text” an AI works with when it reads your writing and produces an answer. Both the amount a model can handle at once and the price you pay are measured in this unit. But one character does not equal one token. That is why counting characters alone will not give you an accurate read on limits or costs.

How is a token different from a character?

Think of tokens as small boxes that a piece of text gets divided into. One box might hold a single character, or part of a word, or an entire word. Even a single punctuation mark can be a token. The box is just a metaphor to help the idea land; the point is that the chunks are not always the same size.

People normally measure the length of a text in characters or words. An AI processes input and output in tokens. So for the same piece of text, the character count a person tallies and the token count an AI processes will not match. That is why “how many characters did I write” and “how many tokens did I use” need to be treated as separate questions.

This distinction is especially important when you give a model a long document. The length of the text you send is a poor guide to how much the AI will actually process. And you have to account not only for what you send, but also for the answer the AI will generate.

How many tokens does Korean text use?

There is no universal formula for converting a Korean character count into a token count. Token counts vary by language, and they also vary depending on which model you use and how the text gets split into chunks. Even if you have seen guidance that estimates tokens from an English character count, you should not apply it directly to Korean.

The official documentation from OpenAI, Anthropic, and Google does not by itself support the claim that Korean always consumes some fixed multiple of the tokens English does. If you pick a multiplier and use it to calculate the length of a Korean document, your figure may diverge from actual usage. If you want to know how many tokens a given amount of Korean text comes to, you need to check against the specific model you plan to use.

For Korean-language users, the important thing is not ranking languages against each other. The point is to avoid applying rough estimates based on English directly to Korean-language work. Two requests with identical character counts can still differ in actual token count and cost.

Does the answer count toward the amount handled at once?

The amount a model can handle in a single request is called the context window. Put simply, it is the total budget available for reading and responding in one go. When you size up that budget, thinking only about the text you just pasted in is not enough.

Prior turns of conversation and your new input both count, and so does the answer the AI produces. So it is a mistake to read a model’s stated token capacity as the amount you can devote entirely to a new document. If you are adding documents over the course of a long conversation, you have to count the earlier exchanges, the new document, and the response together.

This budget differs from model to model. You cannot assume that the limit for one model also applies to another. And since tokens and characters do not line up, statements like “AI can generally take up to this many characters” are not accurate either.

How do you read an AI price list?

As of September 2026, OpenAI, Anthropic, and Google all publish pricing split into input tokens and output tokens. These are the rates that apply when a developer builds AI into their own service. Input is what you send to the model; output is what the model generates. Some models go through an internal reasoning process before producing an answer. That portion, which never appears on screen, is also counted as output and factors into both cost and limits. That is why you need to look at both line items when reading a price list.

Consider two tasks: sending a long text and getting a short answer back, versus sending a short request and getting a long answer back. The cost structures of those two jobs are different. That is why you have to count the tokens in what you send and what you receive separately. The length of your question alone cannot tell you the total cost.

Actual unit prices also depend on the model you choose and whether caching is in play. Caching is a way of reusing previously processed content to reduce cost. Rather than pulling a single rate out of a price list in isolation, it is more accurate to check which model and which conditions that figure applies to. When you compare prices, check first that you are comparing like with like.

What should you check first in practice?

Work through it in three steps. First, decide which model you are going to use. The amount handled at once and the pricing both differ by model. Do not start calculating from the company name alone; reading the documentation for the specific model is the starting point.

Second, break down what will go into the job. For limits, count the prior conversation, the new input, and the answer that will come back. For cost, separate input from output. Splitting things this way reduces the risk of judging both limits and cost from “how long my text is” alone.

Third, check that model’s limits and pricing conditions. Do not multiply a Korean character count by an English conversion ratio and treat the result as settled. When you look at token counts, look at the model and the language; when you look at unit prices, look at input, output, and caching conditions together.

Four things to remember before you start

• Token count is not the same as character count or word count.
• Do not apply English-based estimates directly to Korean.
• The amount handled at once includes prior conversation and the answer.
• For pricing, check the model, input and output, and caching conditions together.

Understanding tokens lets you read limits on long texts and AI usage fees far more accurately. You do not need to chop up every piece of text and count it by hand. Just remember that character counts and token counts are different things. When you pick a model or look at a price list, checking what that model can handle alongside the conditions of your own work is enough.