The price of AI capability is collapsing but the price of using AI is rising: model discipline in the business case

The price of a given level of AI capability is collapsing. The price of using AI is going up. Both are true, and the difference matters for every budget.

This week OpenAI halved the tokens in its US$200-a-month plan and introduced a US$500 tier. Semafor reads it as a sign of a severe compute bottleneck: always-on agents are expensive to run, and there are not enough tokens to make them perform at full potential. OpenAI reportedly spent around US$10 million over a weekend on part of the Navier-Stokes problem.

Meanwhile AI boosters point out, fairly, that last year's capability now costs a small fraction of what it did. The trap is assuming that saving reaches the customer. It does not, because nobody uses last year's model. Usage grows and moves to the frontier faster than unit prices fall.

Any finance team that has run cloud infrastructure has seen this film. Cheaper compute did not shrink cloud bills. It made people do more with it.

The practical response is model discipline: match the model to the task, cap consumption, and stop paying frontier prices for work a smaller model handles well.

If your AI business case assumes costs will fall, check whether it assumes your usage will stay still.

Source: Semafor Technology, "The hand of AI" (30 September 2026), https://lnkd.in/gUkP-WGX

#AI #CostManagement #EnterpriseAI