REPORTED DEVELOPMENT
Shandong researchers Yangze Liu and Zhongyi Han's Harness Tokens study compares Claude Code, miniSWE and OpenCode with fixed models and tasks. Costs differed by roughly three times at equal model prices. Across 447 tasks and two models, Claude Code and miniSWE outcomes were within five percentage points; on 45 difficult tasks, harness switching or repeated runs produced a median 13% outcome difference.
Context, system instructions, tools and step structure drive consumption. Findings are not universal; code and data licensing differ, the dataset does not contain every trajectory and independent replication is absent.
WHY IT MATTERS TO THE MARKET
Model price alone does not determine execution economics. Coding benchmarks do not establish fashion-sector ROI.