Study Finds Code Review Consumes Most Tokens in LLM‑Driven Software Projects
Researchers analyzed token use in a GPT‑5 powered multi‑agent software development framework, mapping stages like design, coding, and testing. The iterative code‑review phase accounts for roughly 59 % of token consumption, with input tokens making up 54 % overall, highlighting hidden costs in automated refinement. The findings guide cost prediction and workflow optimization for AI‑assisted development.