Token spending per engineer is rising with every model upgrade, and that is the signal hiding inside the price war
As AI model prices fall, the cost per engineer is going up, not down. Three data points from inside companies building and using these tools point to the same underlying dynamic: cheaper tokens produce more consumption, not less.
The headline story about AI pricing is that it keeps falling. The less visible story is what happens to spending when it does. Across three separate conversations on separate podcasts, speakers at companies building AI infrastructure and products described the same pattern: as models improve and prices drop, the amount engineers consume rises faster than the price fell. Token spend per engineer is going up, not down.
The clearest articulation came from Cat Wu on Lenny’s Podcast. Wu described a pattern her team observes with each model upgrade or substantial product improvement: people delegate more tasks to the model, spend more hours inside the tool, and the token cost per engineer or knowledge worker rises as a result. The mechanism she described is not complicated. Better models earn more trust, and more trust earns more use. The cost curve follows the capability curve upward even as the per-token price moves the other direction.
Andrew Feldman, speaking on No Priors, supplied the most concrete number. At Cerebras, token spending per engineer went from under one thousand dollars to somewhere between twenty-five and thirty thousand dollars over eight months. That is a thirty-fold increase in less than a year, at a company that builds AI hardware and has both the technical fluency and the financial motivation to understand what it is buying. The figure is internal and unaudited, as all self-reported spending figures are, but it is also specific enough to carry weight.
We lowered the price of it, but the consumption went up way, way more than what you would have expected. Krishna Rao · Invest Like The Best
Krishna Rao on Invest Like The Best named the dynamic directly. When Anthropic lowered the price of one of its models, consumption increased by far more than the reduction in price would predict under standard demand assumptions. Rao called it Jevons paradox, the counterintuitive outcome where efficiency improvements in a resource lead to greater total consumption, not less, because the lower cost makes previously uneconomical uses viable. The paradox is well-documented in energy economics. It is now appearing in AI token markets.
The three accounts are independent. Wu works on a product and is describing user behavior at the application layer. Feldman is describing his own company’s internal spend. Rao is describing what Anthropic observed when it changed its own pricing. None of them are citing each other. All three arrived at the same shape.
What the pattern suggests is that the current framing of AI costs as a procurement problem, something to be managed by selecting the cheapest capable model, may be missing the more durable dynamic. If every capability improvement triggers a new round of expanded use that raises per-engineer spend regardless of per-token price, then the cost question is not primarily about unit economics. It is about how organizations decide what to delegate and at what point that delegation stabilizes, if it stabilizes at all.
None of the speakers claimed to know where the ceiling is. Feldman’s thirty-fold increase over eight months describes a trend, not a terminal state. Wu’s observation about task delegation is behavioral and ongoing. Rao’s Jevons observation describes a response to a single pricing event, not a prediction about the next one. The convergence in their accounts is the signal worth tracking: three different vantage points, three different companies, the same direction of travel.