Tokens and the context window

After this lesson you’ll know: what a token is, why Dutch is more expensive than English, what a context window is, and how to make sure an AI keeps track of things.

What is a token?

A language model reads tokens: common chunks of text. A short, frequently used word is usually one token; a long or rare word gets chopped into pieces. Go ahead and type something below:

Type a sentenceI asked the chatbot for a summary of the annual report.

0 tokens · 0 characters

Approximation: every model splits slightly differently. For exact counts use your provider’s official tokenizer.

• OpenAI: 1 token is about 4 characters, or three-quarters of a word, in English text. Google gives a similar ratio for Gemini.
• Claude: according to Anthropic, about 555,000 words fit into 1 million tokens, just over half a word per token. Claude therefore chops the same text into more tokens.

Pro tip

Only the provider’s own count is reliable. OpenAI has a free online tokenizer at platform.openai.com/tokenizer; Anthropic and Google offer a counting feature in their API (free at Anthropic).

Why Dutch costs more tokens than English

Many tokenizers are tuned to mostly English text, so other languages are often chopped into more pieces. Research from 2023 found differences of up to a factor of 15 between languages, and therefore in price too.

Our own measurement: the Universal Declaration of Human Rights in four languages, counted with OpenAI’s open-source tokenizer.

Language	Tokens	Compared with English
English	1,581	1.00×
German	1,995	1.26×
French	2,049	1.30×
Dutch	2,169	1.37×

Our own measurement, 1 November 2026, using the o200k_base tokenizer (used since GPT-4o). Text up to and including Article 24; a single formal text, so treat it as an indication. With the older GPT-4 tokenizer, Dutch was as much as 1.81× as expensive. On 2,009 more everyday sentences the gap is 24%: see the token index for five tokenizers.

Rule of thumb: for Dutch text with OpenAI models, count on roughly one and a half tokens per word. With Claude, which counts more tokens, it’s higher.

The context window: an AI’s working memory

The context window is all the text a model can “see” at one moment when producing an answer (Anthropic: working memory, Google: reading capacity). Everything counts towards it:

flowchart LR
  subgraph V["Context window"]
    direction LR
    A["Hidden instructions<br/>from the app"] --- B["Memory &<br/>files"] --- C["Your entire<br/>conversation"] --- D["Search and<br/>tool results"] --- E["The answer<br/>it's writing now"]
  end

When the window is full, some apps drop the oldest messages and others automatically summarise older parts. Details from the beginning can then get lost.

Big numbers, small print

“1 million tokens” is usually the technical maximum via the API. In the chat app it depends on your plan (as of 1 November 2026, according to the makers):

App	Free	Paid
Gemini	32,000 tokens	AI Plus 128,000 · AI Pro and Ultra 1 million
Claude	up to 1 million tokens, varies by model	1 million tokens with the 5.5 models

These limits change regularly; if in doubt, check your app’s pricing page.

Pitfall

More context isn’t automatically better. Research (“Lost in the Middle”, 2023) showed that models make best use of information at the start and end of a long text, and more often miss details in the middle. Newer research (Chroma, 2025) tested 18 models and found a more general problem: the longer the input, the less reliable the answers, regardless of where the information is. Anthropic calls this phenomenon context rot.

How to keep the AI sharp

1. One topic per chat. New topic? New conversation.
2. Most important first (or last). Put your key question and most important instructions at the start or end, not in the middle.
3. Give only what’s needed. Paste the relevant chapter, not the whole 200-page report.
4. Bring a summary along. Is a conversation dragging on too long? Ask for a summary and start a new chat with it.

Why do you pay per token?

Via the API you pay per token, because every token costs computing power. Output (what the model writes) is usually several times more expensive than input (what you put in); thinking tokens count as output.

Worked example with OpenAI’s GPT-6.1 Sol ($2 per million input tokens, $10 per million output tokens, as of 1 November 2026):

• You paste a Dutch document of 5,000 words: about 7,300 tokens, so just over $0.01.
• You ask for a 500-word summary: about 730 tokens, so less than $0.01.

About 2 US cents in total. The exact number of tokens depends on the model’s tokenizer, so treat this as an estimate.

Work it out yourself

Enter your own API usage and compare the monthly cost of all major models:

Requests per dayInput words per request your question + context/documentAnswer words per requestDays per month

0 tokens per month

Model	Per request	Per month

Prices in US dollars as published by the providers (excl. VAT), checked on 2026-11-01. Estimate: real token counts differ per model (Claude, for example, counts more tokens per word). Thinking tokens of reasoning models come on top as output.

Free · runs entirely in your browser · nothing is stored or sent

Cost

In the chat app (free or paid) you don’t pay per token but are subject to usage limits. Long conversations, large files and searches use them up faster.

Reuse pays for itself

Do you keep sending the same long instruction or document via the API? With prompt caching, repeated input is billed at a much lower rate. Put the fixed part (instructions, document) first and the question last: cheaper, and the question sits at the end, where it is picked up best.

Security

• Everything in the context window goes to the provider, including attachments, search results and connected email. Paste only the relevant part, without names or numbers (what not to share).
• Text in your context can contain instructions. A researcher got ChatGPT with a Gmail connection to leak emails via a calendar invite. Only connect email and calendar if you really use them (prompt injection).
• An API key never belongs in a prompt or shared chat. Leaked? Revoke it immediately and create a new one (managing secrets).

Checklist

0/4

• I know what a token is
• I know that Dutch costs more tokens than English
• I know what a context window is and what counts towards it
• I start a new chat for each topic

Your progress is only stored in this browser.

Test yourself

1. 1What counts towards the context window?

   Only your last questionEverything: the app's instructions, the whole conversation, files, search results and the answerOnly the files you upload

2. 2Where in a long text does a model most often miss details?

   At the startIn the middleAt the end

3. 3You've been chatting about a project for hours and the answers are getting sloppier. What's the smartest move?

   Ask the same question louder, in capitalsAsk for a summary and start a new chat with itAdd even more text

Next lesson: model names decoded. Why is one model called Sol and another Opus?