What is a token in AI? The unit models bill in

Pieces of text, not words

A language model never sees your words, only the tokens they are cut into. Sometimes a token is a whole word, sometimes part of a word, a punctuation mark or a space with the word after it. A model like ChatGPT or Claude, trained with machine learning, never sees letters or words: it sees a sequence of tokens, computes with them and returns tokens. The provider's invoice counts tokens too, separately for what you send (input) and what the model writes (output).

OpenAI's rule of thumb: one token is about four characters, or three quarters of an English word. The word tokenization, for example, becomes token plus ization. For most European languages that rule is too generous, and it shows on the bill.

How text falls into pieces

The tokenizer is the part that cuts text into tokens. It holds a fixed vocabulary of pieces built from vast amounts of text: what occurs often gets its own token, what is rare gets assembled from smaller pieces. That approach, subword units based on byte pair encoding, comes from a paper by Rico Sennrich, Barry Haddow and Alexandra Birch from August 2015, written for machine translation. The problem they solved was rare words and compounds, exactly what Dutch and German are full of.

We ran OpenAI's public tokenizer o200k_base over a sentence from a transport company. In Dutch, 'Onze vrachtwagenchauffeurs krijgen hun weekplanning automatisch.' is 6 words and 11 tokens: the compound for truck drivers alone breaks into four pieces. The English sentence is 8 words and 9 tokens. Every provider uses its own tokenizer, so the same text weighs differently everywhere. Anthropic itself states that Claude's tokenizer since version 4.7 produces about 30 percent more tokens for the same text.

The language tax: why English is cheapest

A tokenizer's vocabulary is built mostly from English text, so English words fit in a token more often. A study of 25 European languages across ten models, published in May 2026, measured how many tokens a word costs on average. If you run customer service or content in several EU languages, this is the multiplier on your bill.

The same text in German also fills the model's memory faster than in English. For a chatbot on a multilingual website, budget per language, not per conversation.

How much a model reads at once

The context window is the number of tokens a model can handle in one go: your instruction, the attached documents, the conversation so far and the answer together. Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 have a window of 1 million tokens and write at most 128,000 tokens back. Anthropic puts 1 million tokens at roughly 555,000 English words. OpenAI's GPT-6 Luna has a window of 1,050,000 tokens, of which at most 922,000 input.

A large window is no reason to send everything. Long prompts cost more and sometimes a higher rate: GPT-6.1 Sol charges 4 instead of 2 dollars per million above 272,000 tokens, Claude Haiku 5.5 charges 0.50 instead of 0.10 dollars above 100,000. Send only the passages that matter, as RAG does. How to phrase a request so the model needs little is covered under prompt.

Prices per million, per model

Providers charge per million tokens, in dollars, with separate prices for input and output. Output costs five times more for every model below, because writing takes more computation than reading. Standard-tier prices:

These are API prices, for software you have built. A subscription such as ChatGPT Plus is a flat monthly fee: see ChatGPT pricing. Which model fits which task is compared in ChatGPT vs Claude.

1,000 emails summarised: the bill

Say an accounting firm in the Netherlands receives 1,000 client emails a month in Dutch and wants a three-sentence summary of each in the client file. Per email you send a fixed instruction of 150 words plus the email of 250 words: 400 words, at 1.67 tokens per Dutch word about 668 tokens of input. The summary is 60 words, about 100 tokens of output. For 1,000 emails that is 668,000 tokens in and 100,000 out.

On GPT-6.1 Sol that costs 1.34 dollars of input and 1 dollar of output, 2.34 dollars in all, or 2.08 euros at the ECB rate of 21 October. On GPT-6 Luna it is about 10 cents. For Claude, count roughly 30 percent more tokens: 2.71 euros on Sonnet 5.5, 14 cents on Haiku 5.5. The same emails in English cost 1.53 euros on GPT-6.1 Sol; in German, 2.19 euros.

Now set the labour against it: an employee who spends two minutes reading and summarising each email works 33 hours a month on it. Tokens are rarely the problem. The money goes into connecting the mailbox and the client file, testing on your own email and reviewing, the work of AI automation. Run your own volume through the AI cost calculator.

Keeping the bill small

• Pick the smallest model that does the job. Summarising and sorting often work on Haiku 5.5 or GPT-6 Luna; save Sonnet or Opus for real reasoning.
• Put the fixed instruction first and let the provider cache it. A cached block costs 0.10 dollars per million tokens on Claude Sonnet 5.5 and GPT-6.1 Sol instead of 2 dollars.
• Send what is not urgent as a batch: OpenAI and Anthropic both take half off.
• Cap the length of the answer. Output is the expensive side.
• Do not send a whole email thread when the last message is enough.