One thing that seems under-discussed is how quickly token cost changes product behavior once people move from demos to recurring workflows.

When the interaction is exploratory, the marginal cost feels invisible: ask again, summarize again, try another agent. In a business workflow, the same pattern becomes a metering problem. You have to decide which parts actually need a frontier model, which can use a smaller/local model, and which should not be generated at all.

That probably pushes AI products away from "chat with everything" and toward much narrower tools with explicit ROI: less open-ended generation, more constrained pipelines, caching, evaluation, and human review at the points where mistakes are expensive.