What is a context window?

A context window is all the text a model can look at while it writes a reply, and that includes the reply itself. Anthropic's documentation describes it as a kind of "working memory" for the model.¹

It's a space with a fixed size. Once it's full, something has to give.

What goes inside

Many people picture the window as holding only their latest question. It holds much more. Anthropic lists what counts toward it in its developer documentation: the system prompt, every message in the conversation, including tool results, images and documents, and the definitions of any tools.¹

A system prompt is a set of standing instructions that a business or an app maker gives the model before you type anything. You usually can't see it, but it takes up room.

Put simply, the window can contain:

• Hidden instructions from the app you're using.
• Every earlier message in the chat, yours and the model's.
• Any document, image or PDF you've attached.
• Your newest message.
• The reply being written right now.

All of that is measured in tokens, the chunks that models read.

It isn't the same as training

The documentation also draws a line between the window and the large body of data the model was trained on. Training is how the model learned. The window is what it can see right now.¹

That matters for business use. A model doesn't automatically know your price list, your policies or last week's customer emails. If it should use them, they have to be in the window.

How big is it?

It depends on the model. Anthropic's page, checked on 10 November 2026, says many of its current models have a 1M-token window and that other models, including one marked deprecated, have a 200k-token window.¹

OpenAI's help page gives no single figure. It says to check your model's documentation, because the limits can differ between models.²

So the honest answer to "how big?" is: look up the exact model on its maker's page. Sizes change as new models appear.

A bigger window isn't a better memory

It's tempting to think more room always means better results. Anthropic says otherwise. Its documentation states that as the token count grows, accuracy and recall degrade, an effect it calls context rot. It adds that choosing what goes in matters as much as how much space is free.¹

In everyday terms, a stuffed window can make the model less reliable at picking out the one detail you care about. A shorter, tidier prompt often helps. OpenAI's page suggests removing unnecessary or repeated context when input is too large.²

Two limits, not one

The window limits the tokens a model can work with in a request.² Your input and the model's output share that space. Models also have a separate cap on how much they'll write in one go. The page on input and output limits covers that.

What it means for you

Think of the window as the model's desk. Only what's on the desk can be used. Old papers get pushed off, and the desk has a size.

• Put the facts the model needs into the conversation. Don't assume it already has them.
• Leave out the clutter.
• Check the real size for the model you use, on the maker's own page.
• Passwords, card numbers and customer files don't belong on the desk unless your workplace policy says they may be there.

Why a chat can lose track of something you said earlier is the next question, and it has its own page.