A year ago, feeding a long document to an AI meant chopping it into pieces and hoping it remembered the first part by the time it read the last. That constraint is quietly falling away. Newer models — including Claude's Sonnet 5, which now ships with a million-token context window by default — can hold the equivalent of hundreds of pages in mind at once. In plain terms: you can hand the AI the whole contract, the entire project file, the full manual, and ask questions across all of it.

What "context window" even means

Think of the context window as the AI's short-term memory for a single conversation — everything it can "see" at one time, measured in tokens (roughly three-quarters of a word each). A million tokens is somewhere around 750,000 words, or a very tall stack of documents. When the window was small, you had to summarize and feed things in chunks. When it's this large, you can drop the raw material in and let the model work across the whole thing.

What that changes for real work

This is the difference between "summarize this page" and "read these forty files and tell me every place the delivery date is inconsistent." You can load a full vendor agreement and ask what's unusual. You can hand it a year of project notes and ask what slipped. For businesses whose knowledge is scattered across long documents — contracts, specs, policies, case histories — the practical win is being able to ask questions of the entire pile instead of the one page you remembered to paste.

The catch nobody advertises

Bigger isn't free or foolproof. More tokens cost more money, and models still get "lost in the middle" — details buried deep in a huge document can get less attention than what's near the start or end. A million-token window is an invitation to be lazy, and laziness with a contract is expensive. The fix is old-fashioned: point the model at the specific sections that matter, and verify anything you'd act on.

The takeaway

Treat the big context window as a genuine capability, not a reason to stop thinking. Next time you'd normally excerpt a document, try handing over the whole thing and asking a sharp, specific question. Then check the answer against the source. Used that way, a model that reads your whole filing cabinet is one of the most useful tools you're probably not using yet.