A million tokens sounds like you can stop choosing. Paste the whole wiki, the whole codebase, the whole chat, and let the model sort it out. Anthropic's own docs warn against exactly this: as the token count grows, accuracy and recall degrade. They call it context rot, and their engineers say it shows up across all models, not just weak ones.
Why would more text make a model worse? Recall attention from What is an LLM, really?: at each step, a token looks back and decides how much weight to give every earlier token. Anthropic describes this as an attention budget. Every token you add draws on it, and with n tokens there are n² pairs to weigh. The detail you care about has to compete with everything else you sent.
Here's a toy version. It starts on the final “?”: see where that token's attention goes. The weights are illustrative, made up to show the idea.
“?” looks most at “4471” (34%), “code” (18%), “code” (10%).
The answer, “4471”, still gets the biggest share, but a lot of the weight leaks onto lunch, lifts and parking. Three filler sentences are nothing. Now picture three hundred pages of them. That's not a cliff where the model suddenly fails: it's a slope. The model stays capable, but gets less precise about finding and linking details as the context grows.
And at the actual limit? If your input alone is over the window, the API refuses the request with a “prompt is too long” error. If the input fits but the reply runs into the limit, Claude stops early with the stop reason model_context_window_exceeded. The API never silently drops old messages for you. Some chat apps do that behind the scenes; with the API, trimming is your job (or a feature you switch on, which you'll meet in two lessons).
There's a money side too. On Claude Haiku 5.5, a prompt over 100,000 tokens is billed at $0.50 per million input tokens instead of $0.10 (October 2026 prices). Big context costs more per token as well as more tokens.