AI token explained simply: what tokens are and why they set prices and limits

What a token is

An AI does not read text letter by letter, nor word by word. It first splits it into building blocks called tokens. According to OpenAI, a token can be a single character, part of a word, a whole word or a punctuation mark. Spaces and capitalisation also affect how a text is split.

Google describes it in a similar way: a token can be a single letter or a whole short word. Long words are split into several tokens.

The AI processes these building blocks and produces its answer token by token as well. That is why almost everything in AI is counted in tokens: what you type, what the AI answers and how much fits into a conversation.

How many words is a token?

There is no fixed conversion. The vendors only give rough rules of thumb for English text.

OpenAI says one token is about four characters or roughly three quarters of a word, and 100 tokens are about 75 words. Google also gives about four characters per token for its Gemini models, and 60 to 80 English words per 100 tokens.

Both vendors stress that other languages can have a different ratio. The sources reviewed here give no separate figure for German, so we do not give one either. Even the same text produces different token counts depending on the model: according to Anthropic, newer Claude models from version 4.7 onwards split the same text into about 30 per cent more tokens than older models.

Why tokens set the price

If you use AI models directly through the vendors' programming interface, for example as a developer or company, you pay per token. Prices are quoted per million tokens, in US dollars and without any statement on VAT. Input and output are priced differently, and output is usually much more expensive. Some vendors also charge more when the input is very long.

Some examples, as of October 2026: OpenAI charges 2 dollars per million input tokens and 10 dollars per million output tokens for its GPT-6.1 Sol model on the standard tier with short context; with long context it is 4 and 15 dollars. Anthropic charges the same 2 and 10 dollars for Claude Sonnet 5.5. Google charges 0.75 dollars for input and 3.75 dollars for output for Gemini 3.8 Flash until 31 December 2026; from 1 January 2027 these prices double, according to Google.

Thinking is a hidden cost factor: with so-called reasoning models, OpenAI says internal thinking tokens are generated that you do not see as answer text, but that are billed as output. Google also counts thinking tokens towards the output price. A short visible answer can therefore cost more than its text suggests.

What tokens mean for you on a chat subscription

If you simply use ChatGPT, Gemini or Claude in the app or browser, you do not pay per token but use it for free or with a monthly subscription. Tokens still matter in two places.

First, for the length of a conversation. How much text a single chat can hold is measured in tokens and is called the context window. In the free version of ChatGPT, OpenAI states 27,000 tokens for quick answers; for Gemini without a subscription, Google states 32,000 tokens. We have a separate explainer on the context window.

Second, for usage limits. Anthropic writes that for Claude, the length and complexity of your conversations, the model you choose and the effort level all affect how quickly you reach your limit. For Gemini, Google says its limits also take into account the complexity of your prompt, the models and features used and the length of the chat.

How to save tokens in everyday use

A few simple rules can be drawn from the vendors' advice. Paste only the part of a document you actually need, rather than the whole file. Start a new chat for a new topic instead of continuing a very long thread. OpenAI advises removing unnecessary or repeated content and splitting large texts into smaller parts.

For Claude, Anthropic also recommends turning off thinking and any tools or connected apps a task does not need, and choosing a lower effort level for routine questions. According to Anthropic, tools use particularly many tokens.