Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering (arxiv.org) AI

The arXiv paper analyzes token consumption in an LLM-based multi-agent system for software engineering, using ChatDev execution traces across SDLC stages (design, coding, code completion, code review, testing, and documentation) to quantify input/output/reasoning costs. It reports that iterative code review is the dominant source of tokens (about 59.4%), and that input tokens make up the largest share overall (about 53.9%), suggesting that refinement and verification drive much of the expense rather than initial code generation.