Context window
How much input a model can see at once. Bigger is partly true, mostly misleading, and always expensive.
The context window is the amount of input (measured in tokens, not pages) that a model can consider in a single call. A 200,000-token window means the model can take in about 150,000 words of text at once. Past that limit, earlier material gets dropped or summarized away.
In 2024, common windows were 8K to 32K. In 2026, frontier models routinely accept 200K to 1M+. The headline number has grown roughly 50× in two years.
Why it matters now
A larger window changes what’s possible. You can drop an entire codebase, a quarter of earnings transcripts, or a thousand-page deal binder into a single prompt and ask for analysis across all of it. Workflows that used to require chunking, retrieval, and stitching can sometimes collapse into one call.
Sometimes — but not always. Two facts the spec sheets bury.
First, attention quality degrades inside the window. A model with a 1M-token window doesn’t pay equal attention to all 1M tokens. The first 30% and the last 10% reliably get the most attention; the middle, less. “Lost in the middle” is the term of art.
Second, every token costs. A million-token call is a million tokens billed. At current frontier prices, dropping a single deal binder into the model can cost more than the analyst hour it replaces.
The misuse
The vendor framing is “longer context = more capable.” The operator framing should be “what’s the smallest input that gets the right output reliably?” Long context is a fallback, not a default. Retrieval-augmented designs that pull the relevant 5,000 tokens beat 500,000-token brute-force prompts on both quality and cost in most enterprise tasks. Use the long window for the cases retrieval can’t handle, not because the spec sheet says you can.