What is a context window, and why do long conversations get worse?

What actually fills the context window

Every message you send and every reply the model gives accumulates inside the window as a conversation continues — earlier turns aren't summarized away on their own, they're preserved in full.

• The system prompt
• Every message in the conversation, including tool results, images and documents
• The definitions of every tool made available to the model
• The output generated in that turn, including extended thinking

Why more context isn't automatically better

This is the part that surprises most people, and it's the reason this page exists: a bigger window is not simply a bigger advantage.

The practical consequence is that curating what's in context matters just as much as how much space is available. A window with plenty of room to spare can still be full of the wrong thing.

How big it is, and what counts against it

Size depends on the model. Claude Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5 and Sonnet 4.6 all have a 1M-token context window on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry — as do Claude Fable 5 and Mythos 5. Other models, including Claude Sonnet 4.5, have 200k.

In the 1M-token models, a single request can generate up to 128k tokens of output, and a request can include up to 600 images or PDF pages — 100 in the 200k models.

Every response reports exactly how much of the window it used, in a usage field. With prompt caching on, that input count splits into input_tokens, cache_read_input_tokens and cache_creation_input_tokens — and all three still count toward the window. Caching changes what those tokens cost you, not whether they take up space.

What happens when you go over it, and how it's managed for you

If the input alone is already bigger than the window, the API rejects the request outright with a 400 invalid_request_error saying the prompt is too long. In Claude 4.5 models and newer, it's less blunt: if input plus the max_tokens you asked for would exceed the window, the request is accepted anyway, and generation only stops if it actually reaches the limit, ending with stop_reason: “model_context_window_exceeded”.

There's also a token counting API for estimating how much of the window a request will use before you send it. And three things happen to manage the window for you, without being asked:

• Server-side compaction — in beta for Claude 4.6 and newer models — summarizes the older parts of a long conversation on Anthropic's server, and is the primary strategy for keeping one going past what would otherwise fit.
• Context editing offers tool result clearing and thinking block clearing — removing specific things from context instead of summarizing everything.
• Context awareness, in Claude Sonnet 5, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5, lets the model track its own remaining budget — the API injects that information automatically, so there's nothing to turn on.