Token Price
The token price indicates what a provider charges for processing a certain amount of text chunks through an AI model. It is the standard billing unit for language AI and determines whether a product can be operated economically.
Programs like ChatGPT break down every text into small building blocks. Such a building block is usually a short word or part of a word, referred to in technical jargon as a token. The word “umbrella” can, for example, split into “um” and “brella”. The token price is the amount a provider charges for processing such building blocks. Because a single building block is extremely cheap, prices are almost always quoted per one million building blocks. Typical figures range, depending on the model, from a few cents to several dollars per million.
What the price reveals about an AI business
Anyone who integrates an AI into a product doesn’t pay once, but with every use. A chatbot with one million conversations per day generates costs every day. The token price is thus the single most important cost lever. If it halves, the room for new features doubles.
This explains why price cuts by AI providers are treated as major news. Between 2023 and 2025, prices for comparable performance fell by more than a factor of ten. Applications that were previously unaffordable suddenly became realistic. These include, for example, programs that automatically summarize entire mountains of files.
A common misconception: the lowest price is not automatically the cheapest. A weak model often needs several attempts or longer answers for the same task. In the end, what counts is the price per solved task, not per building block. This is precisely what experts argue about when comparing models with one another.
How the bill is composed
Providers distinguish two kinds of building blocks. Input building blocks are everything you send in: your question, attached documents, the conversation history so far. Output building blocks are what the model writes. Output typically costs three to five times as much as input.
The reason lies in the technology. The model can process the input in parallel in a single pass. The answer, on the other hand, is generated block by block, each requiring its own computation step. More computation steps mean more power consumption and more time on expensive graphics chips.
A short calculation makes this tangible. Suppose input costs one dollar per million building blocks and output costs four dollars. A request with 2,000 building blocks of input and 500 building blocks of response then costs around 0.004 dollars. At one million requests per day, that amounts to about 4,000 dollars daily. An additional trick is called caching: recurring text segments are cached and then billed at a heavily reduced rate.
Token prices in news and products
The token price affects you most directly if you program yourself. Anyone using a provider’s interface, i.e. a technical access point to the model, will see a pricing table there. It lists two figures for each model: input and output per million building blocks. Monthly subscriptions like ChatGPT Plus hide these costs behind a flat fee.
In business news, token prices appear as a signal for competition. When one provider lowers prices, others must respond. Analysts then speak of a price war, similar to mobile phone tariffs. At the same time, investors ask whether providers even make a profit at these prices.
The token price should be distinguished from context length. This indicates how many building blocks a model can take into account at once. A large window allows for long documents but makes individual requests expensive. And beware of confusion: the token on a blockchain is something completely different from a token in language AI.
