Liu and Han evaluated five models with Claude Code, OpenCode, and mini-SWE-agent, including repeated runs on SWE-bench Verified. For two models on the larger pool of 447 tasks, Claude Code and mini-SWE-agent were equivalent within five percentage points, while OpenCode trailed by up to nine points. Per-task costs for the same model differed by as much as threefold, largely because system prompts and tool definitions recur in the input. On the 45-task hard subset, both switching harnesses and simply rerunning a setup flipped 13% of outcomes, limiting what a small one-off score difference can establish.