Jesse Zhang, the Suna AI cofounder, posted on X that a unit of inference ran about $60 per million tokens in 2020 and now costs pennies. Despite that collapse, he argues, enterprise AI spending keeps clearing budget lines set when tokens were expensive. The gap is not a pricing story. It is a consumption story.
This is the Jevons paradox showing up in inference economics. William Jevons observed in 1865 that more efficient coal-burning engines did not reduce coal consumption; they made coal-powered work cheap enough that total usage rose. Cheaper tokens have the same effect on language models. When a task costs a fraction of a cent, teams stop rationing it, and usage expands faster than the per-unit price falls.
Reasoning models accelerate that expansion. A model that thinks before answering can burn tens of thousands of tokens on chain-of-thought reasoning that never appears in the final response the user reads. Agentic workflows compound the effect further: an agent that plans a task, calls tools, checks its own output, and retries on failure can multiply a single user request into dozens of model calls. None of that shows up in a per-token price sheet. It shows up in the invoice.
Zhang’s framing puts responsibility on the buyer side, not the vendor side. Falling list prices create a natural assumption that costs will fall in step. Finance teams built AI budgets on 2023 and 2024 usage patterns, when a single prompt meant a single completion. Agentic and reasoning products broke that assumption without changing the line item finance teams were tracking.
The practical failure is measurement, not the pricing model. Teams that track dollars per user or dollars per seat miss the driver entirely, because token volume per task is what moved. A support bot that answered in one pass last year and now runs a five-step agentic loop this year can cost more per resolved ticket even at a lower per-token rate. Multiply that pattern across every internal tool a company rolled out over the past two years, and small per-task increases add up to a budget line finance did not model.
Operators should treat token consumption as a metered utility, not a fixed software license. That means instrumenting cost per completed task, not just per API call, and setting alerts on token volume growth separately from dollar spend, since volume is the leading indicator and the invoice is the lagging one. Any team evaluating a reasoning model or an agentic framework this quarter should benchmark tokens consumed per successful task alongside accuracy, not after it.
Jesse Zhang, on X, July 20, 2026.
