A token is the basic unit of text that an AI model reads and writes, and it is roughly a word-piece rather than a whole word — which is why AI tools bill per token instead of per word, counting the tokens you send in (input) and the tokens the model sends back (output) separately.
If you have ever looked at an AI bill and wondered why a long document cost more than a short question, the token is the answer. It is the meter on the taxi: distance is what you pay for, and the token is the distance.
So what actually is a token? When you send text to a model, the model does not see letters or words the way you do. A piece of software called a tokenizer chops your text into chunks and assigns each chunk a number.
Common words often become a single token. Longer or rarer words get split into several. Punctuation, spaces, and even the way you format a list all become tokens.
As a rough rule of thumb, English text tends to run somewhere around four characters per token, so a paragraph of a few hundred words might be a few hundred tokens. That is a rule of thumb, not a guarantee — the exact count depends on the tokenizer, which differs between vendors.
This matters because everything downstream is measured in tokens: how much the model can hold in its context window, how fast it responds, and what it costs you.
Here is a worked example you can actually picture. Suppose you paste a 2,000-word contract into an AI tool and ask, "Summarize this in five bullet points." Your input is the contract plus your instruction, so call it a couple of thousand tokens in.
The model's output is five short bullets, maybe a hundred tokens out. Now compare that to a different task: you ask the same tool to rewrite the entire contract in plain English. Your input is similar, but your output is now thousands of tokens instead of a hundred.
Same document, wildly different bill, because output tokens are the expensive half. This is the single most useful thing to understand about AI pricing: input and output are counted separately, and output is usually the pricier side. If you want to cut costs, shorten what you send and ask for shorter answers — not just shorter prompts.
Where this advice runs into limits. Token counts are not stable across models or languages. A tokenizer built for English handles English efficiently and handles other languages poorly, so the same sentence in a language the tokenizer was not optimized for can cost noticeably more tokens.
Code, URLs, and long strings of numbers also tokenize badly — a random ID string can eat far more tokens than its character count suggests. On top of that, exact per-token prices change often, and they differ between models and between input and output. Our AI tool database tracks a pricing and capability snapshot for 360 AI tools, with the most recent verification date of 2026-09-18, but even that is a snapshot rather than a live feed.
For anything you are about to spend real money on, the vendor's own pricing page is the only reliable source, because it is the only place updated the moment a price changes. A final tip that surprises most beginners: you are often billed for tokens you did not consciously send. System prompts, tool definitions, and previous conversation history all count as input on every turn, so a long chat can quietly get more expensive as it goes.
Starting a fresh conversation when you switch topics is one of the simplest ways to keep the meter from running.