Inference got 430× cheaper. Your agent didn't.
Inference got 430× cheaper per token between 2021 and 2026, but a resident agent burns a hundred times more tokens than a 2021 chatbot query. Do the division and the headline savings shrink from 430× to 4.3×; run agents as hard as I do, into the millions of tokens a day, and they vanish entirely. It's the Jevons paradox on the agentic frontier: cheaper inference didn't become a discount, it became a budget we spent on capability. The per-token price chart and the per-task cost curve point in opposite directions, and pricing a business off the wrong one is how the margin math breaks.