token-based pricing

in English: per-token pricing

basicwidespread

A pricing scheme where the AI provider bills based on the amount of text processed, meaning the number of input and generated tokens.

In plain terms

Language models measure text not in letters or words but in tokens: short text segments, such as a word, part of a word or a punctuation mark. With token-based pricing, you pay for as many tokens as the model read and wrote for you.

It is like an electricity bill: most of the cost depends on the kilowatt-hours consumed. A short question typically costs less than one forint, but processing many long documents or thousands of conversations becomes a noticeable expense.

This is different from a monthly subscription, where you can use the chatbot as much as you want for a fixed sum (within certain limits). Token-based fees are typically paid by companies and developers who integrate the model into their own programs, websites, or agents.

How it works

The provider breaks the submitted text (question, instructions, attached documents, previous conversation) into tokens and counts them: these are the input tokens.
The model's response also consists of tokens: these are the output tokens. Their unit price is typically a multiple of the input tokens because the model processes the input all at once, while the response must be generated token by token sequentially, which is slower and ties up more machine time.
Prices are usually quoted per one million tokens, varying by model: larger, more powerful models are more expensive, while smaller, faster ones are cheaper.
The bill is the sum of the two: input tokens × input unit price, plus output tokens × output unit price. The "thinking" tokens of reasoning models, which are internal and mostly hidden from the user, also count as output.
Many providers offer discounts if your requests start with the same long text (caching) or if the response is not urgent and requests can be processed in batches.

An everyday example

An online store runs AI customer service: for every customer question, the model receives the store's policies, the conversation history, and the new question, then answers. At the end of the month, the store does not pay "for one chatbot," but for all the tokens consumed across the conversations – more at the Christmas peak, less in January.

Who is affected

developerThey must design the application so that it does not send unnecessarily long texts to the model, because every token costs money. → Code

entrepreneur, executiveAI costs are not a fixed monthly fee but increase with usage, so consumption must be calculated in advance and monitored.

employeeIf your company's AI tools are billed by token, your usage habits show up in the costs, and a cap may be set on them. → Work & HR

A common misconception

MythA token is a word, so by counting words I can know exactly how much I will pay.

In realityA token is often just a word fragment. Hungarian text typically splits into more tokens than the same text in English, because tokenizers were mainly trained on English, and inflected words and accented letters break down into more parts. Additionally, the bill includes instructions and context sent in the background.

Limits and risks

The cost is hard to predict in advance: the length of the response and the model's thinking time vary from request to request.
Agents work in multiple steps, sending the entire history with each step, so a single task can consume far more tokens than expected.
Providers' tokens are not uniform: the same text yields different token counts across different models, so unit prices alone are not comparable.
A cheap token does not mean a cheap solution: if a weaker model makes more mistakes and needs reruns, the final total may be higher.

Similar, but not the same

subscription pricing – There you pay a fixed monthly fee largely independent of usage, while here you are billed separately for each processed text segment.
token – The token itself is the unit of measurement (a piece of text), and token-based pricing is the fee structure built on this.
outcome-based pricing – You pay for the completed task or achieved result (for example, a resolved customer complaint), not for the amount of text consumed in the process.

Related concepts

Broader

usage-based pricing

Narrower

bemeneti token, kimeneti token, caching, batch processing

On the edge

token, tokenization, language model, API, context window, subscription pricing, inference cost

A brief history

Per-token billing became known through the application programming interface (API) of OpenAI's GPT-3 model. The API launched privately in 2020 and initially offered packages with a monthly token quota. From 2021, usage-based pricing quoted per thousand tokens became the norm. Since then, nearly all major providers have adopted the method. As prices have fallen significantly, the usual unit for quoting prices has become one million tokens.

Where things stand

By the end of 2026, this is the default pricing for programmatic access to language models. With the spread of agents, task- and outcome-based pricing experiments have also emerged, but on the provider side, these are typically underpinned by token costs.

Status: 2026 November

Also known as

token-based pricing · token-based pricing · per-token pricing · token price · per-token pricing · token-based pricing · pay-per-token · price per million tokens

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.