Study Reveals Code Review Devours 60% of Tokens in AI-Driven Software Development

New research analyzing GPT-5 multi-agent software development traces reveals that automated code review devours nearly 60% of total token consumption across the SDLC, while input tokens make up the majority of usage. The findings challenge assumptions that AI coding costs concentrate in initial generation, instead highlighting iterative refinement and verification as the primary expense. Researchers mapped ChatDev execution traces to standardize evaluation across design, coding, testing, and documentation phases, offering a framework to predict costs and optimize agent collaboration protocols.