The short answer
A token is the small chunk of text, often part of a word, that an AI language model reads and writes. Anthropic's glossary describes tokens as the smallest individual units of a language model, which can be words, subwords, characters or bytes. Google's rule of thumb for Gemini is that one token is about 4 characters and 100 tokens is about 60 to 80 English words. Context windows, usage limits and API prices are all counted in tokens.
What a token is, in plain English
Before a large language model can read your sentence, it splits the text into tokens. A common short word is often one token. A long or unusual word may be split into several pieces. Spaces, punctuation and numbers also become tokens. The model then predicts the next token, then the next, which is why answers appear word by word.
Anthropic's glossary describes tokens as the smallest individual units of a language model, corresponding to words, subwords, characters or even bytes. It explains the trade-off: larger tokens make the model more efficient, while smaller tokens let it handle rare or never-seen words. Each model family has its own tokeniser, so the same text produces a different count in ChatGPT, Claude and Gemini.
How many words is a token?
There is no single answer, but the makers give rules of thumb.
- Google (Gemini): a token is about 4 characters, and 100 tokens is about 60 to 80 English words.
- Anthropic (Claude): on earlier models a token is about 3.5 English characters. Claude 4.7 and later models use a newer tokeniser that produces about 30 percent more tokens for the same text, so a newer model counts the same document higher.
- Language matters. Anthropic notes that characters per token depend on the model and the language. English usually tokenises most efficiently; many other languages use more tokens for the same meaning.
A useful working estimate: a page of plain English text is a few hundred tokens, and OpenAI's pricing page lists a 27K-token window for GPT Instant on ChatGPT Free, with room for about 12 pages of text as input.
Images, audio and video are tokens too
Google's Gemini documentation spells out how other media are counted: an image up to 384 pixels on both sides is 258 tokens, larger images are cut into 768 by 768 tiles of 258 tokens each, video is 263 tokens per second and audio is 32 tokens per second. A 10-minute recording is therefore about 19,200 tokens of audio alone. See what is multimodal AI.
Why tokens matter at work
You rarely see tokens in a chat app, but they decide three things you do notice.
- How much fits. A model's context window is measured in tokens. Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 take up to 1M tokens in the Claude API, and a single request can generate up to 128k output tokens.
- How fast you hit limits. Long files, long chats and agent runs burn through plan allowances because they use more tokens. That is why a big PDF can exhaust a free plan's daily limit faster than dozens of short questions.
- What you pay. Developer APIs charge per token, and Anthropic bills a reasoning model's thinking as output tokens. Tools built on top of these APIs increasingly pass that on. Since 1 June 2026, GitHub Copilot plans bill in GitHub AI Credits based on tokens at each model's API rate, with annual Pro and Pro+ subscribers moving over when their plan ends. Bolt's free plan gives 1M tokens a month with a 300K daily limit. Cursor bills third-party models such as Claude Opus 5.5, Claude Sonnet 5.5 and Gemini 4 Argon separately by usage, at API list price from prepaid credit, on Premium and higher plans.
Try it: count the tokens in your own text
- Take a real document from your week, such as a two-page brief.
- Count its words in your word processor. Multiply by about 1.3 to 1.7 for a rough English token estimate based on Google's rule of thumb.
- Now paste the same text in a different language, or a page of code or a spreadsheet export, and notice that you cannot rely on the same ratio.
- For exact counts, developers can use the token counting tools in the makers' APIs: Anthropic's token counting endpoint for Claude and the count tokens method in the Gemini API.
Once you can estimate, you can plan. If a report is around 60,000 tokens, you know it will not fit in ChatGPT Free's 27K-token GPT Instant window, so split it or use a tool built for long documents. See how to summarise a PDF with AI.
Common misconceptions about tokens
- "A token is a word." Sometimes, but often a token is part of a word, a space plus a word, or a punctuation mark.
- "Token counts are the same across tools." Each model family uses its own tokeniser, and Anthropic says its newer tokeniser produces about 30 percent more tokens for the same text than earlier Claude models.
- "Only my message counts." The model's reply, the conversation history, attached files, instructions and, on reasoning models, the hidden thinking all count. Anthropic bills thinking tokens as output tokens.
- "AI tokens are crypto tokens." They are unrelated. An AI token is a unit of text, not a digital asset.
- "Fewer tokens means a worse answer." Trimming irrelevant material usually improves answers, because recall drops as the context window fills.
Related terms
- Context window is the token limit for one request.
- Large language model is the model that reads and writes tokens.
- Inference means running the model to generate tokens, which is what you pay for.
- Embedding means turning text into numbers so it can be searched by meaning.
- AI API is where token pricing is most visible.
Prompts to copy
Fit a long document into a small window
I will paste a long document in parts. For each part, write a summary of no more than 150 words that keeps every figure, date and name. Do not answer any questions yet. After the last part I will type DONE, and then answer: [your question].
You get: A chunked summary that fits a small context window, followed by an answer based on all the parts.
Ask for a shorter answer that keeps the facts
Rewrite your last answer in half the length. Keep every number, name, date and action point. Remove repetition and background I already know.
You get: A tighter answer that uses fewer tokens when you paste it into the next step.