Study: coding-agent harness adds no score gain but shifts cost up to 3x

Researchers at Shandong University compared Claude Code, OpenCode and mini-SWE-agent on the same models across SWE-bench Verified tasks. Score differences stayed within noise, while cost per task varied up to 3x: 1.81 CNY for Claude Code versus 0.56 for OpenCode on GLM-5.3-Flash.

• Claude Code's first call sends 17,134 tokens; mini-SWE-agent sends 1,382
• On a 45-task hard subset, switching harnesses flipped about 13% of results
• OpenCode trailed by up to 9 points, largely from not re-prompting after output-cap cutoffs
• On DeepSeek-V4-Flash, mini-SWE-agent made 141 model calls per trial versus 104 for Claude Code