What you'll pay for AI agents will be wildly variable and unpredictable

ZDNET summarizes a University of Michigan-led study showing that agentic AI systems (agents) consume orders of magnitude more tokens than turn-by-turn prompt interactions, exhibit large run-to-run variance, systematically underestimate token needs, and are dominated by input/cache-read token costs — producing unpredictable and potentially high monetary bills for users and enterprises.