Claude Code Infrastructure Cuts Token Burn 68%

Hot take: this is an environment problem masquerading as a model problem.

• –The strongest signal here is not the exact 68.5% number, but the shape of the win: fewer text hops, fewer parse steps, fewer re-queries.
• –Human-oriented tooling like shell output, grep, and cat forces agents to spend tokens reconstructing state that could be exposed directly.
• –The claim is compelling, but it reads like a single benchmark rather than a broad controlled study, so the exact savings will vary by workload.
• –If you run agents at scale, structured state APIs, semantic retrieval, and less chatty tool surfaces are probably the first cost lever to audit.