Cost spreads 17.5x across nine harnesses on same model

Kimi K3 ran through nine coding-agent harnesses in 12 configurations, 360 runs total on 30 identical tasks, per a benchmark posted to frontier-harness-eval/eval.

Same model every time. Only the harness changed. Cost per passing run ranged from $1.05 to $18.34, a 17.5x spread that falls entirely on whoever pays the token bill for that harness.

Codex passed 66.7% of tasks at a median $3.47 per pass. Claude Code matched DSH Creator's 63.3% pass rate but cost $18.34 per pass, more than 5x Codex's price for a lower pass rate. Exo Harness passed 53.3% at $1.05, the cheapest in the set.

Cache behavior accounts for most of that spread. Claude Code's median cache hit rate was 67.8%, well below the field. Every point missed means the agent re-sends context it already paid to process once.

The repo doesn't say why Claude Code's cache hit rate lags the field, only that it does.