What Are Tokens in AI? A Plain-English Guide to How Chatbots Count Text (and Bill You)

TL;DR: Tokens in AI are the small chunks of text that language models read and write. A token can be a whole word, part of a word, a single character or a punctuation mark. In English, one token is roughly four characters or about three-quarters of a word, so 100 tokens is about 75 words, according to OpenAI. Tokens matter for two practical reasons: they set how much a model can handle at once (its context window), and they’re how AI APIs charge you, usually as a price per million input and output tokens.

If you’ve ever wondered what tokens in AI actually are, you’re not alone. The word appears on every pricing page, in every model launch (“1 million-token context!”) and in error messages when a prompt is too long. Yet chatbots rarely explain it. This guide covers what a token is, how text gets split into tokens, the different kinds of tokens you get billed for, and how to estimate token counts and costs without doing any maths you don’t need to.

What is a token, exactly?

OpenAI’s help article on tokens describes them as “the units that OpenAI models use to process text,” noting that a token “can represent a character, part of a word, a whole word, or punctuation.” Models don’t see letters or words the way we do. Before your prompt reaches the model, a tokenizer breaks it into pieces from a fixed vocabulary and turns each piece into a number. The model works entirely with those numbers, then converts its output numbers back into text.

Common words usually get a single token. Rare words, names, typos and long technical terms are often split into several. Spaces and capitalization matter too: OpenAI points out that “red”, “Red” and ” red” (with a leading space) can be tokenized differently.

Why not just use words or letters?

It’s a trade-off. A vocabulary of every possible word would be enormous and would still miss new slang and names. Working letter by letter would make every sentence very long for the model to process. Sub-word tokens fall in between. A model can handle almost any text, including made-up words and code, by combining familiar pieces. Many tokenizers build their vocabulary with a method called byte pair encoding, which repeatedly merges the most common character pairs in training text into new tokens.

Quick conversions: tokens, words and characters

OpenAI gives these rules of thumb for English text. They’re estimates, not exact counts:

Two warnings. First, other languages often use more tokens for the same meaning. OpenAI notes that “other languages can have different relationships between characters, words, and tokens.” Second, different companies’ tokenizers split text differently, so the same paragraph can count as more tokens on one model than another.

The four kinds of tokens you’ll see on a bill

Not every token is priced the same. OpenAI’s documentation separates them like this:

• Input tokens (also called prompt tokens): everything you send, including your question, any pasted documents, system instructions and chat history.
• Output tokens (completion tokens): the text the model writes back. These usually cost more per token than input.
• Cached input tokens: repeated prompt text the system has already processed recently, often billed at a discount.
• Reasoning tokens: internal “thinking” that reasoning models do before answering. OpenAI says these aren’t shown as answer text but “count toward output usage and are billed as output tokens.” A short reply can therefore use far more tokens than it appears to.

Context windows: the token budget per conversation

Every model has a context window, the maximum number of tokens it can consider at once. That includes your prompt, attached files, the conversation so far and the answer it’s writing. It also has a separate cap on output length. When a long chat starts to “forget” early details, or an upload is rejected as too large, you’ve usually hit that limit.

Context windows have grown fast. For example, Inception’s new Mercury 3 model advertises a 1 million-token window, up from 260,000 for its previous version, according to the company. By the table above, that’s roughly 750,000 words of English. Bigger isn’t automatically better, though. You pay for every input token you send, so stuffing a huge context into every request gets expensive.

How token pricing works (with a worked example)

AI APIs typically quote prices per million tokens, with separate rates for input and output. Using Mercury 3’s list price of $0.20 per million input tokens and $0.70 per million output tokens as an example:

1. You paste a 1,500-word article and ask for a summary. That’s about 2,000 input tokens.
2. The model writes a 300-word summary, about 400 output tokens.
3. Cost: 2,000 × $0.20 / 1,000,000 = $0.0004 input, plus 400 × $0.70 / 1,000,000 = $0.00028 output. The total is well under a tenth of a cent.

Individual requests are cheap, but costs scale with volume. An app making that request 100,000 times a month would spend about $68 at those rates. OpenAI also warns that a lower price per million tokens “does not necessarily produce a lower total cost,” because models tokenize text differently and may generate more output or reasoning to finish the same job. For a real-world comparison of how one company tiers its models, see our guide to GPT-6 Sol vs Luna.

How to count tokens yourself

• Online tokenizer: OpenAI offers a Tokenizer tool that shows exactly how a piece of text is split, with each token highlighted.
• In code: OpenAI’s open-source tiktoken library counts tokens for its models. Pick the right encoding with tiktoken.encoding_for_model(model).
• After the fact: API responses report usage, as prompt_tokens and completion_tokens in Chat Completions or input_tokens and output_tokens in the Responses API.
• Quick mental maths: word count ÷ 0.75 gives a decent estimate for English.

Tips to use fewer tokens

• Trim the context. Paste only the relevant section of a document, not the whole file.
• Start new chats for new topics. Long threads resend history, which adds input tokens on every turn.
• Ask for the length you need. “Answer in three bullet points” saves output tokens, which are usually the pricier kind.
• Write clearer prompts. A precise instruction avoids back-and-forth. Our plain-English guide to prompt engineering covers how.
• Reuse fixed instructions. If your app sends the same long system prompt every time, prompt caching (where supported) can make those repeated tokens cheaper.

The bottom line

Tokens are the currency of language models. They decide how much a model can read, how long it can write and what you pay. You don’t need to count them precisely to use chatbots well. But knowing that 100 tokens is about 75 English words, that output costs more than input, and that reasoning models quietly spend extra tokens will help you understand pricing pages, avoid context-limit errors and keep API bills predictable.

About Teck Hustlers Editorial Team

The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.