Tokenomics study finds 59.4% of costs in AI code review, not generation

A new research paper titled "Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering" (arXiv:2601.14470) by Mohamad Salim and colleagues provides the first empirical breakdown of token consumption in LLM-based multi-agent (LLM-MA) systems for software development. Using ChatDev—a framework that orchestrates multiple GPT-5 agents across the entire software development lifecycle—the team analyzed execution traces from 30 tasks, mapping internal phases to standard stages: Design, Coding, Code Completion, Code Review, Testing, and Documentation.

The study’s core finding is that the iterative Code Review stage accounts for a staggering 59.4% of total token consumption on average. Furthermore, input tokens (prompts, context, and conversation history reused across agents) represent 53.9% of all tokens, suggesting severe inefficiencies in agent collaboration—agents repeatedly feed each other the same information. This challenges the common assumption that initial code generation is the primary cost driver; instead, the real expense lies in automated refinement and verification loops. The authors provide a standardized framework for practitioners to predict and optimize token budgets, and they call for research into more token-efficient protocols for multi-agent systems.