How the Cost Climbed Up
Unit prices are clearly falling, so why is total cost rising? The article breaks out the mechanism: agentic workflows are token-devouring beasts—within a single task, the model repeatedly reads context, calls tools, and self-corrects, consuming tens to hundreds of times what a single Q&A does; the more complex the task, the more iteration rounds, and failed retries are all money too. Multiplied out, an agent doing serious work already racks up an hourly bill that can go toe-to-toe with human hourly wages in some regions—and it still makes mistakes that a human has to clean up.
Both Sides of the Equation Are Moving
The value of this piece is in pulling "automation necessarily saves money" back from faith to arithmetic. Of course, both sides of the equation are changing: engineering optimization of inference cost is ongoing, and hardware iteration keeps pushing prices down; but on the other side, vendors' pricing has to recoup astronomical training investments, so the pressure to raise prices is just as real. The pragmatic move is to run the unit economics before starting an agent project—which tasks' "AI hourly wage" is really below a human's, and which just look cool. Do the math clearly before you launch, so you don't agentify everything only to find the savings don't even cover the bill.
via: Hacker News