AI context windows explained

When you ask an AI model to summarize a report or answer questions about a long conversation, it needs space to process the text. Its context window is the maximum amount of information it can handle at once, measured in tokens, or small units of text. Even within that limit, the model can miss details.

The context window holds information the model can use while answering. It does not, by itself, store information between chats. Whether information carries over depends on how the product stores and supplies it.

What counts toward an AI context window

A request can contain more than the words you type. Its context may include instructions, conversation history, documents you provide, tool results, and the model’s response. Depending on the model and product, tokens used for internal reasoning also count toward the limit, even if they do not appear in the answer.

This is why a long conversation can eventually push earlier messages out of reach or leave less room for a response. Models count input and output in different ways: some use one shared limit, while others set separate limits. A product can also impose additional restrictions on request size or response length.

Tokens are not words

Models measure text in tokens, small units that can represent a whole word, part of a word, punctuation, or other text. A passage’s token count depends on how the model encodes text, as well as the language and wording.

OpenAI gives a rough estimate of about three-quarters of an English word per token. Use it as a rule of thumb, not an exact conversion. A page or word count cannot tell you exactly how much of the context window a document will use. This is especially true for requests that include images or files, and for those that use tool calls or structured data.

AI context window limits by product

The context-window limit listed for a model does not always tell you how much text you can submit. Check whether the figure applies to the model and product you use, and whether it covers only input or both input and output.

Documentation checked on October 29, 2026, gives these examples:

These figures describe different models and product setups, so they do not provide a fair comparison of model quality. They also do not show how much useful input a particular request can hold once you include instructions, conversation history, space for the response, and other limits.

A larger context window does not guarantee recall

Even when a model accepts a long document, it may overlook a detail inside it. In the 2024 study Lost in the Middle, researchers tested how models found information across multiple documents. For the models and tasks they tested, results depended on where the relevant information appeared. Models often did better when the information appeared near the beginning or end than when it sat in the middle.

That study does not establish how every current model handles every kind of request. It does show why context capacity and reliable use of context are different measures. Anthropic’s prompting guidance similarly recommends placing long documents before the question in certain large-document tasks. That advice comes from Anthropic’s own guidance and tests. It does not apply as a universal rule for every model.

Give long prompts more structure

For a long report or bundle of documents, organize the information so the model can find what matters. Group related material and give sections descriptive headings. Also state what the model should find or compare. If a specific passage matters, identify it directly rather than assuming the model will notice it in a large collection.

When accuracy matters, test the model you plan to use with documents and questions like the ones you expect to send. Check whether it finds details in different parts of the input, and ask it to point to the supporting passage if the product allows it. If it misses key information, shorten or reorganize the material and try again. A larger context window lets more text fit in one request. The window’s size alone does not show how well the model will handle that material.