Input Token
Input tokens are the small text chunks into which an AI program breaks down the text you send it. Their number determines how much a request costs and how much text the program can even take in.
A chat program like ChatGPT doesn’t read text letter by letter or word by word. It first breaks it down into small chunks. These chunks are called tokens. A token is often a short word, sometimes just a syllable or a punctuation mark. Everything you send in yourself — your question, an inserted text, an instruction — becomes what are called input tokens. Everything the program writes back, on the other hand, counts as output and is counted separately.
The unit in which AI is billed
Anyone who doesn’t use AI through an app but instead builds it directly into their own software pays per token. Providers state prices per one million tokens. Input tokens are almost always cheaper than output tokens, often by a factor of three to five. The reason: generating new text is more computationally intensive than merely reading in existing text.
That sounds like small change, but it adds up quickly. A company that attaches a 30-page manual as background knowledge to every request might send around 20,000 input tokens per question. At a million requests a month, that becomes a real cost factor. That’s why developers shorten their inputs as much as possible.
A second point is the upper limit. Every model has a context window: the maximum number of tokens it can process at once. Input and output must fit in together. If the input is too long, the request fails or older text gets dropped.
From sentence to list of numbers
The breakdown is handled by a program component called a tokenizer. It works with a fixed list of roughly 50,000 to 200,000 known text pieces. Frequent words appear in it as a whole, rare ones are assembled from several pieces. “House” is one token, “Donaudampfschifffahrt” breaks down into several.
Afterwards, each token is given a number from the list. So the model doesn’t compute with letters, but with a sequence of numbers. As a rule of thumb for German: one token corresponds to roughly 0.6 to 0.7 words. German texts need more tokens than English ones, because German words are longer and are frequently compounded.
A common misconception is that input tokens are only one’s own text. In practice, a lot is often added that you don’t see: a system instruction from the provider, the prior conversation history, attached documents. Images are also converted into tokens. A photo can cost several hundred to over a thousand input tokens.
How to recognize input tokens in price lists and news
The pricing tables of OpenAI, Google, or Anthropic always have two columns: input and output. Anyone building an AI feature into a website calculates precisely with these figures. Many providers now offer discounts on repeated inputs, so-called prompt caching. If you send the same long preceding text multiple times, it costs only a fraction the second time.
Tokens also show up in reports about new models. When it’s said that a model has a context window of one million tokens, that means: you can present it with roughly a thick book all at once. Such figures are an important selling point in the competition among providers.
For private users, tokens usually remain invisible, because subscriptions are billed as a flat rate. Still, the limit becomes noticeable: if a very long chat suddenly forgets what was said at the beginning, the context window is full. Older input tokens were then discarded to make room for new ones.
