Same model, up to 5x the bill: FrontierHarness v1.1 tests 22 coding-agent harnesses across 27,280 runs
Runta's new benchmark ran five open models through 22 coding-agent harness setups and found the wrapper can multiply cost per task by up to 5x with little or no gain in pass rate, and that the cheapest harness for one model can be among the priciest for another. Here's what the numbers show and where they stop.