What's in the window

A model knows two kinds of things. What it soaked up in training is baked into its weights: broad, fuzzy, and frozen at a cutoff date. What you put in front of it right now is the context window: the text it can actually read while it writes. Andrej Karpathy's way of putting it is that the weights are a vague recollection and the context is working memory.

Working memory is the part you control. Anthropic counts everything in the request toward it: the system prompt, every earlier message in the chat (including tool results, images and documents), the definitions of any tools you hand it, and then the reply Claude writes, thinking included.

How big is it? As of October 2026, every current Claude model (Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, the model in this sandbox) has a 1M-token window, and one reply can be up to 128k tokens of it. Some older models stop at 200k. A million tokens is several long novels, so on paper you can fit a lot.

Two things it is not. It isn't memory between requests: the API is stateless, so a chat app re-sends the whole conversation every turn, and anything you don't send, Claude has never seen. And it isn't the training data: pasting a document in doesn't teach the model anything for next time.

You can see the statelessness right here. Every run in this sandbox is a fresh request with nothing carried over. Ask Claude who's on call this weekend and it can only guess. Then give it the rota, Saturday: Priya, Sunday: Marco, in the same prompt, and ask again.