The price fell and the bill rose

On Tuesday the two largest American laboratories cut their prices within an hour of each other, and this desk wrote up the supply side of it. On Wednesday the demand side arrived. McKinsey's senior partners Tanguy Catlin and Lari Hämäläinen used a client session titled "Improving the Economics of Agentic AI" to make a point that sits awkwardly beside a price war: the cost of intelligence is collapsing and enterprise AI bills are going up anyway.

The arithmetic of the paradox is not in dispute. GPT-4 arrived in 2023 at $60 per million output tokens. Models that match it on the standard benchmarks now run at a small fraction of that, and Tuesday's cuts moved the frontier price down again — Anthropic's cache reads fell 60 per cent in a single release. "Intelligence at a certain capability level is getting a lot more affordable," Hämäläinen said, as Fortune reported. What has grown faster than the price has fallen is the quantity consumed.

McKinsey's own survey supplies the scale. Fielded between 4 May and 8 June across 1,719 respondents in 97 countries, it finds roughly one in five now say AI-related operating costs — token costs among them — are constraining how much AI they use. About a third put more than a tenth of their technology budget into AI. Sixty per cent expect to spend more next year regardless.

The mechanism is agents, and specifically the fact that an agent decides for itself how much work a task requires. A multistep job can be completed by a short path or a long one, and the firm reports that cost "variance can be up to 30x between different runs" of the same task. Hämäläinen's framing is the sharpest: "Imagine like you're running an operation, but every day there's a 30x difference in the cost."

That is a different complaint from the one usually filed under AI spending. A level can be budgeted; a level that moves thirtyfold for constant output cannot. Procurement is built to negotiate unit prices, and the unit price here is falling. What it is not built to do is forecast volumes set by a model's own judgement about how hard it should try.

The economics driving the volume explain why cheaper models make bills larger rather than smaller. McKinsey's threshold for a viable agent is that a success rate above roughly 10 per cent can generate value provided checking the work costs about a tenth of doing it manually. That is an argument about the cost of verification, not about model quality, and it licenses a strategy of many cheap attempts: every reduction in the price of an attempt widens the set of tasks worth attempting badly. Cheaper cache reads, likewise, reward feeding the model more context rather than less.

Software development is where this lands hardest, because coding agents are, as the firm puts it, very token hungry. It is also where the second-order effect shows up. Nearly a third of organisations report deciding against buying at least one software product or feature because they could now build it in-house with agentic tools. The AI line on the budget is not simply growing; it is partly eating the enterprise software line next to it. A finance chief challenged on AI spend has an answer available that most are not yet giving: some of it is not new money.

Against that sits the return. The same survey finds 37 per cent attributing any EBIT impact at all to AI, unchanged on last year, and 6 per cent qualifying as high performers deriving more than 5 per cent of EBIT from it, also unchanged. Eighty per cent of individual users report being more productive. The gap between the personal experience and the profit-and-loss statement has not closed in twelve months, while the bill has grown.

McKinsey's remedies are three: visibility into which use cases, units, agents and models are driving spend; workflow optimisation, meaning matching model complexity to the task, routing requests, caching reusable context and capping tool calls and agent loops; and sourcing discipline, meaning unused licences removed, quotas managed, terms negotiated and single-vendor dependence avoided. All three are sensible. Two of them are also products. Routing and caching are now features the laboratories sell, which means the most-recommended cost control is one the vendor configures, meters and prices.

The firm's own caution is the right one: cost reduction should not be indiscriminate, since the objective is to spend more on the highest-return applications and less on the rest. But that presupposes the organisation can tell which is which, and the survey's finding that only a third of firms can see where their AI spend is going suggests most cannot. Consumption pricing transfers the forecasting problem from the seller to the buyer. The laboratories know precisely what a token costs them. Their customers, on the present evidence, do not know how many they are about to buy.

The price war is therefore not the relief it looks like from the buyer's chair. It lowers the cost of every individual call and raises the number of calls worth making, and the second effect has so far been the larger one. The discipline question for the next four quarters is not whether models get cheaper — they will — but whether finance can learn to meter something engineering has just made elastic.