The demo was never lying: a few dollars of API spend can draft what might have taken a human afternoon. In production you run two ledgers. One is the vendor bill—frontier models are still on the order of dollars per million input tokens and more for output (OpenAI API pricing). The other is retries, review, security fixes, and escalations when “almost right” ships.
Enterprise demand has not plateaued. Menlo Ventures estimates roughly $37 billion in generative AI spend in 2025, about 3.2× the prior year—so even when per-token rates fall, volume and scope (agents, long contexts) can still inflate invoices.
The Stack Overflow 2025 Developer Survey fills in the labour column: 80% use AI tools, but only 29% trust their accuracy (down from 40%). 45% cited “almost right, but not quite” as the top frustration; 66% spend more time fixing that code; 75% still ask a person when they distrust the model. Tokens plus senior attention is the real line item.
Compare that to loaded engineer cost—design, test, ownership, risk—not headline salary. Fast Company has documented teams where junior roles thinned while seniors absorbed work models could not own. The sensible move is selective: humans where ambiguity and liability dominate; automation where work is bounded and checks pass. That recalibration is what makes human capacity competitive again on the paths that matter.
Who is summing API spend, rework hours, and escalation risk—and calling that the breakeven, not just shipping speed?