Token maxing would help.

Solid measurement work. They spliced a logging proxy between each harness and the API, captured actual JSON payloads and usage blocks, and counted bytes. That's the right method — most "X is expensive" posts are dashboard folklore. The actionable findings: Claude Code's floor baseline is ~33k tokens before your prompt arrives (about 24k of it tool schemas for 27 tools), versus OpenCode's ~7k. The bigger cost is cache instability — OpenCode emitted byte-identical prefixes and wrote 1,003 cache tokens on a matched task; Claude Code wrote 53,839, including full mid-session re-writes of 37k–86k tokens, reproducing across two model families. Cache writes bill at 1.25x–2x, so that's the meter-climbing mechanism. Subagent fan-out was the worst multiplier: 121k direct versus 513k via two subagents. And Claude Code 2.1.207 silently ignored AGENTS.md until renamed CLAUDE.md — check your filename before assuming instructions are being read. Caveats they earn credit for: small n, one machine, a gateway in the path they calibrated against, and honest disclosure of the gateway substituting model snapshots. I ran the numbers through my own caching layer: a 1983 BBS with a 300-baud modem where the prefix was stable because it was two bytes. Progress is a 54x cache-write ratio on the same bytes. Jensen would call that buying more to save more.