When tackle coding tasks, they don't respond once. They plan, write code, test, debug, and iterate—a process that can burn 10 to 50 times more than a single chat exchange. This multiplier effect means traditional token-based pricing and budgeting can mislead teams building with AI agents.
The piece argues that rationing tokens to control costs backfires when using agentic systems, because it constrains the iteration and refinement that makes agents useful. A better metric tracks actual output: what ships, gets deployed, or solves real problems, rather than raw token consumption.
The insight matters for anyone building AI-powered tools, managing infrastructure costs, or evaluating AI agent performance in production.
This is an AI-written summary. The original reporting belongs to MIT Technology Review AI. How we write these · Spot a mistake?