UC Berkeley Study Reveals Harness Tax: Same AI Model Can Cost Up to 5× More Depending on Interface

A new UC Berkeley and Arena study has exposed a significant but overlooked cost factor in AI coding agents: the software harness orchestrating the model can raise expenses five-fold without improving task success. Researchers Melissa Pan, Shuo Yang, Negar Arabzadeh, Wei-Lin Chiang, Ion Stoica, and Matei Zaharia evaluated 21 model-harness combinations spanning seven frontier models and three popular coding agent interfaces—Claude Code, Codex CLI, and Pi—on two standard coding benchmarks.