Tokens and Context Windows Explained

Two terms appear in almost every AI announcement and pricing page: tokens and context window. They explain why a chatbot forgets the start of a long conversation, why a big PDF may be refused, and how developers are billed. Neither is complicated.

What a token is

Models do not read letters or whole words. They read tokens: common chunks of text. A short common word is usually one token. A long or unusual word is split into several. Punctuation and spaces count.

Rules of thumb for English:

• 1 token ≈ 4 characters
• 1 token ≈ ¾ of a word
• 100 tokens ≈ 75 words
• 1,000 tokens ≈ 750 words, about a page and a half

Estimates for English text at about 1.3 tokens per word. Exact counts vary by model and language.

These are averages. Three things push the count up:

• Other languages. Many use more tokens per word than English.
• Code, numbers and unusual names.
• Formatting such as tables and JSON.

Images, audio and video are converted to tokens too, by different rules for each model. You can estimate any text with our token calculator.

What the context window is

The context window is the maximum number of tokens a model can take into account in one go. It is the model's working memory for a conversation.

Everything counts toward it:

• the hidden instructions the app gives the model;
• every message you have sent in the chat;
• every reply it has given;
• files, images and search results it has read;
• the answer it is about to write.

Windows have grown from a few thousand tokens to hundreds of thousands, and in some models a million or more. That is enough for long reports or a sizeable codebase. It is still a limit.

What happens when you hit the limit

Different products handle it differently:

• Older messages are dropped or summarised. The model seems to "forget" what you said at the start.
• The upload is refused or truncated. Only part of a large file is read.
• Quality drops before the limit. Models tend to use information at the beginning and end of a long context better than the middle.

Signs you are near the limit: the assistant contradicts an earlier instruction, repeats itself, or loses track of details.

Practical fixes

• Start a new chat for a new topic. It is the simplest improvement most people can make.
• Ask for a summary, then continue from it. "Summarise what we have decided so far" — and paste that into a fresh chat.
• Give only what is needed. Paste the relevant section instead of the whole document.
• Put the instruction after long material. Ask the question at the end, where it is less likely to be lost.
• Break large jobs into parts.

Input, output and reasoning tokens

• Input tokens: what you send, including history and files.
• Output tokens: what the model writes.
• Reasoning tokens: some models think through a problem before answering. That thinking uses tokens you may not see, and they are usually billed as output.

There is also a separate maximum output limit: the longest single reply a model will produce, which is much smaller than its context window.

How pricing works

Subscriptions (the chat apps) charge a flat monthly fee with usage limits. You do not see tokens, but limits are based on them, which is why long conversations and big files use your allowance faster.

APIs (for developers) charge per token, quoted per million tokens, with separate prices for input and output. Output typically costs several times more than input.

A worked example with illustrative prices of $3 per million input tokens and $15 per million output tokens:

• A request with 2,000 input tokens and 800 output tokens costs (2,000 × $3 + 800 × $15) ÷ 1,000,000 = $0.018.
• 500 such requests a day is $9 a day, about $270 a month.

Real prices vary widely between models and change often; use the provider's pricing page and our calculator to run your own numbers.

Ways developers cut cost

• use a smaller model for simple tasks;
• shorten prompts and trim history;
• cache repeated context, which many providers discount;
• batch non-urgent work, often at a lower rate;
• cap the output length.

Why bigger is not always better

A huge context window invites you to paste everything. Resist it. More text means more cost, slower replies and more chance that the important detail is overlooked. The best results usually come from the smallest context that contains what the model needs.

How this connects to model names

Within one family, larger models generally handle long, complex contexts better and cost more per token; smaller ones are cheaper and faster. That is the trade-off behind names such as Astra and Instant, Pro and Flash, or Opus and Haiku. Our model tracker lists the current families, and GPT-6 explained walks through one of them.

Frequently asked questions

How many words is 1,000 tokens?

About 750 words in English. Other languages, code and numbers usually use more tokens for the same length.

What is a context window in AI?

The maximum amount of text, measured in tokens, that a model can consider at once, including your messages, files and its own reply.

Why does my chatbot forget what I said earlier?

Long conversations can exceed the context window, so older messages are dropped or summarised. Starting a new chat with a short summary helps.

Why do output tokens cost more than input tokens?

Generating text takes more computation than reading it, so providers charge more for output.