Claude Code harness costs 1.8x pi

An analysis of 180 runs across three Claude models and three harnesses found that changing the harness dramatically altered API costs while leaving task success rates largely unchanged. Claude Code proved significantly more expensive than minimal harnesses like pi or opencode due to factors such as larger starting prompts, more round trips, stricter caching, and sandboxing blocks. The author argues that teams paying per token should test lighter harnesses before switching models to reduce expenses. Open article →