The one-line version
A token is a small piece of text. Roughly four characters or three-quarters of a word in English. Every word you send and every word the AI sends back gets counted in tokens.
What a token actually looks like
AI doesn't read whole words the way we do. It chops text into smaller chunks. Common words like "the" or "and" become one token. Longer or unusual words get broken up. For example:
- "cat" = 1 token
- "unbelievable" = 3 tokens (un + believ + able)
- A short sentence like "I love pizza" = about 4 tokens
- A 750-word document = roughly 1,000 tokens
- A 10-page contract = roughly 5,000 tokens
Quick math: take your word count and multiply by 1.3 to get a rough token count.
Why tokens matter to you
Three reasons.
- Pricing. AI companies charge per million tokens. The more you send and the more it generates, the more you pay.
- Limits. Every model has a maximum number of tokens it can handle at once. That is the "context window." Past that limit, it forgets the beginning.
- Speed. More tokens means slower responses. A 50-page document takes longer than a paragraph.
Input tokens vs output tokens
Two flavors. Both get billed but at different rates.
- Input tokens. What you send in. Your question, attached files, instructions, chat history. Usually the cheaper side.
- Output tokens. What the AI generates back. Always more expensive, often 4 to 5 times more. Generating new text is harder for the model than reading existing text.
When AI reads your input, it processes the whole thing at once. When it writes a response, it has to predict each token one at a time, which takes more compute. You pay for that extra work.
What a million tokens actually buys you
To make it real, here is what one million tokens looks like in human terms:
- Roughly 750,000 words
- About 1,500 pages of typical text
- The full text of a long novel
- Around 3,000 emails at average length
So when you see a model priced at "$3 per million input tokens," that is $3 for processing the equivalent of a novel.
Current pricing in 2026 (per million tokens)
| Model | Input | Output | Best for |
|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | Fast, simple tasks |
| Claude Sonnet 4.6 | $3 | $15 | Daily work, writing, analysis |
| Claude Opus 4.7 | $5 | $25 | Hard reasoning, complex docs |
| GPT-5.5 | $5 | $30 | OpenAI's main model |
| GPT-5.5 Pro | $30 | $180 | Heavy reasoning, only when you need it |
Most people will never see these numbers because they pay a flat monthly subscription. Tokens only show up directly if you use the API.
How tokens connect to context windows
The "context window" is the total number of tokens the AI can hold in its head at once. That includes your question, attached files, the conversation history, and the response it's generating. In 2026, top models handle 200,000 to 1,000,000 tokens of context, which is hundreds of pages.
When you hit the limit, the model starts forgetting the earliest parts of your conversation. If a long chat starts feeling forgetful, that's why.
The advanced take: prompt caching
If you keep sending the same big system prompt over and over (which is common for businesses), AI companies offer "prompt caching." The model remembers the cached part so you only pay full price once. After that, the cached portion costs about 90% less with Claude and 50% less with OpenAI.
For a small business doing repeated tasks like contract review or intake summaries, this is the difference between a $200/month AI bill and a $40/month one.