What Is a Context Window?

4 min read

Everything a model can use to answer you has to be in front of it when it answers: your message, the chat so far, any files or instructions the app added. That workspace has a hard size, measured in tokens, called the context window.

The window is not a memory. It is more like a desk. Things on the desk are visible. Things filed away might as well not exist until something puts them back on the desk.

How much a model can consider at once
GPT-3.5 · 2023
16K TOKENS
GPT-4 Turbo · 2023
128K
Claude 3 · 2024
200K
Frontier · 2026
1M+
A whole novel is around 120K tokens. Today's frontier models, GPT-5.5, Claude Sonnet 5, and Gemini 3, all sit around a million, and it is still finite.
If it is not in the window, the model is not considering it.

What actually fills the window

More than you type. The app adds its own standing instructions, your recent messages, retrieved files or search results, and notes it saved about you. A short question can arrive at the model wrapped in thousands of tokens of context you never see.

The parts you cannot see get priority, too. The app’s own instructions load first and almost never get trimmed. That is why an assistant keeps its personality through a marathon chat while your third message quietly disappears. You are sharing the window with the app itself.

Big windows help, then stop helping

Modern windows fit whole books, which is genuinely useful. But models pay uneven attention across a huge window, and important details buried in the middle of a long dump get missed. A smaller amount of relevant context regularly beats a larger amount of noise.

Researchers have also found that models pay the most attention to the beginning and end of a long context and lose track of things in the middle. So if something really matters, put it at the top of your message or repeat it at the end. If it is buried in paragraph twelve of pasted text, expect it to get skimmed past.

The window resets, memory does not

Every conversation starts with a fresh desk, which is why a new chat knows nothing about the last one. Anything that survives across chats survives because the product saved it and put it back. That system is AI memory, coming right after forgetting.

Products that feel continuous are doing real work behind the scenes on every turn: reloading your preferences, pulling up the relevant note, adding the project brief back in. People who build AI products call this context engineering. When an app seems to just know things, someone built that.