
AI coding agents are burning tens of thousands of tokens before they even read your prompt
A new comparison of Claude Code and OpenCode found one harness sends 4.7x more tokens than the other before processing a single instruction — and the gap compounds badly once you add subagents.










