Claude Code sends 33k tokens before reading the prompt, while OpenCode sends 7k tokens

Testing Claude Code and OpenCode under the same model and environment confirmed that Claude Code's token consumption is overwhelmingly higher.

Key comparison results:

• Baseline overhead: For a simple response request, Claude Code uses about 33,000 tokens (including system prompt, tool schemas, etc.), while OpenCode uses only about 7,000 tokens.
• Cache efficiency: OpenCode reduces cost by reusing the request prefix, but Claude Code rewrites tens of thousands of tokens mid-session, consuming up to 54x more cache tokens than OpenCode.
• Configuration and scaling costs: Configuration files like CLAUDE.md add an average of 20,000 tokens per request, and connecting 5 MCP servers adds another 5,000-7,000 tokens. Also, when using sub-agents, token consumption per task spikes by roughly 4x or more (121k → 513k).

Conclusion: Claude Code has the advantage of batching tool calls to reduce overall token usage in multi-step tasks, but its baseline cost is very high. Therefore, when operating agent systems, it is essential to manage the impact of token overhead on cost, latency, and context budget with data.