Study finds Claude Code sends ~33K tokens before reading user prompt, 4.7x more than OpenCode

Short summary

A wire-capture study found Claude Code sends ~32,800 tokens before reading your prompt versus OpenCode's ~6,900—a 4.7x gap driven by 27 tool schemas vs 10. Worse, Claude Code's cache writes were 53.7x higher due to shifting system bytes invalidating prompt cache breakpoints, resulting in 3.7x total token spend for identical work. Practical fixes: audit MCP server count and trim instruction files.

• •Claude Code burns ~32,800 tokens before reading your prompt vs OpenCode's ~6,900—4.7x overhead from tool schemas and system prompts
• •Cache writes were 53.7x higher on Claude Code due to shifting system bytes invalidating cache breakpoints
• •Audit MCP server count (5 servers add ~5-7K tokens/request) and trim instruction files to cut per-request costs

Generated with AI, which can make mistakes.