tl;dr: Measuring at the API boundary with a logging proxy, Claude Code sends ~33k tokens of system prompt, tool schemas, and injected scaffolding per request versus OpenCode's ~7k, and rewrites its prompt cache mid-session up to 54x more often—driving large bills even on identical tasks where both harnesses produce correct output. Real configurations (instruction files, MCP servers, subagents) push baselines to 75k–500k+ tokens before any user input, though Claude Code's aggressive tool-call batching can close the gap on multi-step tasks depending on the model.