AI

Stop counting tokens, start counting what ships

The article argues that companies should measure AI coding agents by the software they ship rather than by token consumption. Agentic coding tasks can use 10 to 50 times more tokens than a basic chat prompt because agents plan, generate code, run tests and retry. It cites Uber reportedly exhausting its 2026 AI budget by April, GitHub Copilot’s move to metered credits and Gartner’s prediction that coding-token costs will exceed average developer salaries by 2028. The author says usage caps can restrict productive work because they do not distinguish waste from software delivery.

AI-written summary, checked against the cited sources. How summaries work

Image: MIT Technology Review AI

Read at MIT Technology Review AI