Claude Code Costs, Act III — The ecosystem of options for spending less

Short summary

Analysis of LLM cost optimization tools and how compression composes with prompt caching in Claude Code. Headroom achieves 10-20% real savings by compressing tool outputs without breaking cache. Key insight: rewriting cached bytes costs more than it saves—preserve cached prefix byte-for-byte while compressing only fresh content.

• •Headroom compresses tool outputs upstream of the cache, preserving the cached prefix and achieving 10-20% actual savings
• •Compression and caching compose only if cached bytes remain unchanged; rewriting them voids the 90% cache discount
• •Three cost surfaces exist: cached input (0.1×), uncached input (1-2×), and output (priciest); Headroom targets output compression

Generated with AI, which can make mistakes.