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

What happened

Systima.ai logged every API payload when comparing Claude Code and OpenCode, finding Claude Code consumes ~33,000 scaffolding tokens before reading a user prompt, versus ~7,000 for OpenCode — a 4.7x baseline gap.

Context and impact

Cache instability in Claude Code caused up to 54x more cache token writes on identical tasks. Real-world config multipliers (large CLAUDE.md files, MCP servers, subagent delegation) can push costs 4x higher still. The analysis became the day's top HN post with 421 points, driving broad developer discussion about agent harness cost efficiency.

Details

• Claude Code: ~33,000 scaffolding tokens before prompt; OpenCode: ~7,000 (4.7x ratio)
• Cache instability: up to 54x more cache token writes on identical tasks
• CLAUDE.md, MCP servers, and subagents can further 4x costs
• HN: 421 points — one of the top AI threads on July 12, 2026