Why a bigger context window is not simply better
How the Models Work · July 2026
AI can now read a whole book in one go. It pays least attention to the middle.
The word for this is everywhere now: context window. It is how much a model can read at once: your prompt, the conversation, the documents you paste in. In two years it went from a few hundred pages to thousands, and some models now take ten thousand pages or more in a single go.
But bigger is not simply better. The more text a model reads, the less attention any single fact gets. This is called dilution, and it touches the whole document, not just one part. And the middle suffers most, an effect with its own name in the research, lost in the middle: recall is strongest for material near the start and end of what the model reads, and weakest in between (first documented at Stanford in 2023, replicated since).
How well the model recalls a fact, by position and context length
Longer context dilutes every position; the middle drops most.
Three practical habits follow. First, front-load what matters: put the key facts or the question at the top or bottom of what you paste, never buried in the middle. Second, trust the advertised number less: "reads a million tokens" rarely means it uses them all well, and the reliable range is often 60 to 70 percent of the advertised size. Third, give it less: the strongest fix for dilution is not pasting everything, but extracting or summarising what matters before you hand it over.
Bigger context windows genuinely help. The quiet lesson is that dilution sets the floor, and placement decides who feels it least. Manage both, deliberately.