What Is a Token in AI? A Plain Explanation

A token in AI is the small chunk of text a language model actually reads and pays for, roughly four characters of English or about three-quarters of a word. This paragraph alone runs about 45 tokens. Every model, from GPT-5.6 to Claude, breaks your prompt and its own reply into these pieces before doing anything else, and the token count is exactly what shows up on your bill.

Why models chop text into tokens instead of words

Language models don't read whole words the way people do. They run text through a tokenizer that splits it into sub-word pieces using a compression method called byte-pair encoding, which learns the most common chunks of characters in a huge body of text and assigns each one an ID number. Common English words like "the" or "and" usually become one token. Longer or rarer words, brand names, and most non-English text get split into two, three, or more pieces.

That's why "token" and "word" are not interchangeable, even though people use them loosely in conversation. "Tokenization" itself, for example, splits into "Token" and "ization", two tokens for one word. A model never sees the sentence as a string. It sees a list of numbers, feeds them through billions of parameters, and predicts the next token, one at a time, until it decides to stop.

How much text is actually one token

The commonly cited rule of thumb, OpenAI's own help documentation, is that one token is about 4 characters or 0.75 words of English prose, so 100 tokens is roughly 75 words. Code, non-English languages, and text with lots of punctuation or unusual spelling tokenize less efficiently, sometimes using 2 to 3 times more tokens for the same amount of visible text.

Tokenizers also aren't identical across vendors or even across a single vendor's model generations. Anthropic notes that Claude 4.7 and later models switched to a newer tokenizer that produces roughly 30% more tokens for the same input text than Claude 4.6 and earlier used, which changes real-world cost even before you factor in per-token price changes.

What tokens actually cost: five common tasks priced out

Token counts only matter because they're the billing unit. Below are rough token estimates for five everyday tasks, priced against two current models: OpenAI's mid-tier GPT-5.6 Terra and Anthropic's Claude Sonnet 4.6, using OpenAI's published API pricing and Anthropic's published API pricing as of this writing. These follow the same recent OpenAI price cut that lowered Terra and Luna tier rates in late July 2026. Before running a task like this yourself, it helps to estimate the tokens for an AI task first, so you know roughly what it will cost before you run it.

GPT-5.6 Terra is priced at $2.00 per million input tokens and $12.00 per million output tokens; Claude Sonnet 4.6 runs $3.00 and $15.00 per million respectively. Output tokens cost far more than input tokens on both, which is why a task that generates a lot of text, like rewriting a code file, costs more per run than one that mostly reads text, like summarizing a transcript, even when the transcript has more total tokens.

Why the token count still matters beyond cost

Every model also has a hard ceiling on how many tokens it can hold in a single request, both your input and its output combined. That ceiling is the model's context window, and it's a separate constraint from price. A cheap model with a small context window can still fail outright on a task, like feeding it a 300-page contract, that a pricier model handles in one pass.

This is also why techniques like retrieval-augmented generation exist: instead of stuffing an entire knowledge base into the prompt and paying for every token of it on every request, the system retrieves only the relevant passages first, then sends just those tokens to the model. Standardized approaches like Model Context Protocol also affect token counts, since every tool definition and connected data source a model can see adds tokens to the input on every single call, whether or not that tool actually gets used.

Practical ways to cut your token bill

• Trim system prompts and tool definitions. Every token in a system prompt or tool schema is billed on every request, even ones that never use that tool.

• Cache repeated context. Prompt caching on both OpenAI and Anthropic APIs charges a fraction of the standard input rate for content that was already sent recently.

• Match the model to the task. A short email reply doesn't need a flagship-tier model; a cheaper tier costs a fraction as much for output that's functionally identical for simple text.

• Summarize before you send. Passing a condensed version of a long document instead of the raw text can cut input tokens by an order of magnitude.

Frequently asked questions

How many tokens is 1,000 words?

Using the roughly 0.75 words-per-token rule of thumb, 1,000 words of typical English prose comes out to about 1,300 to 1,400 tokens. Text with unusual formatting, code, or non-English languages will run higher.

Are tokens counted the same way for GPT and Claude?

No. Each model family uses its own tokenizer, and even different generations from the same company can tokenize the same text differently, as with Claude's 4.7-and-later tokenizer producing about 30% more tokens than its earlier models for identical input. There's no universal token count for a given piece of text.

Do you pay for input tokens, output tokens, or both?

Both, but at different rates. Every model API charges separately for input tokens (your prompt, any attached files, and conversation history) and output tokens (what the model generates), and output is consistently priced several times higher than input across every major provider.

Does a longer prompt always cost more?

Yes, a longer prompt always adds more input tokens and therefore more cost, though input tokens are the cheaper half of the bill. The bigger cost driver is usually how much text the model generates in response, since output pricing is typically 5 to 6 times the input rate.

Can images and audio be tokens too?

Yes. Multimodal models convert images, audio, and video into tokens as well, using a separate calculation based on resolution or duration rather than character count, so a single image can add anywhere from a few hundred to a few thousand tokens depending on its size.

Tokens are specific to how language models process text. Image generators work on a different mechanism entirely, covered in what is a diffusion model.

For turning this into an actual cost estimate before you run a task, see how to estimate tokens for an AI task.