Dive into FrontierHarness Eval: Claude Code cost 5.6× for the same pass rate

Given the [K3 setup](https://runta.com/blog/introducing-frontierharness-eval/), how much of the cost gap comes from Claude Code’s harness design versus its fit with the model and caching backend?

Claude Code is an opinionated harness. I’d love to see whether neutral harnesses can challenge the monolithic harness narrative.

Co-author here. Failed attempts still cost money, which is easy to miss when comparing pass rates.