Million-token AI is sold by the window. The work is paid by the read.

Frontier AI is sold by the window size. The work is paid by the read.

When a coding agent opens a real document, most of its compute is spent before the answer: gathering context, comparing passages, deciding which snippet to weigh. Marketing sells the input ceiling. Bills get paid in retrieval rather than reasoning.

The reusable mechanism is attention cost. Standard attention scales quadratically: doubling the document quadruples the compute, and almost all of that compute is paid before the model produces a single token of answer. A team that budgets an "AI project" the way it budgets a model upgrade is budgeting the wrong lever; the constraint to plan around is context cost, not context window.

Alexander Whedon, on the Sept. 8 Eye on AI appearance as Subquadratic's CTO, put a number on the asymmetry: 86% of what frontier coding agents do is "read steps," and accuracy on financial documents at 500,000 tokens drops below 50%. A million-token window is not a million-token working unit.

The 86% is Whedon's framing from one podcast, and a vendor-neutral long-context benchmark may move it. The cost shape it points to is structural, and that survives a correction.