Coding Agent Harnesses Differ in Token Use

The New Stack reported that Aider, Claude Code, and OpenClaw each ran an identical model yet showed token use varying 70-fold. The account states that agent harnesses can change coding-agent costs by multiples even when the model stays fixed. Three benchmarks are cited as showing where the extra tokens come from. The post appears on the account focused on at-scale software development, deployment, and management, and links to an article on the same topic.