Aider, Claude Code, and OpenClaw ran an identical model. Token use varied 70-fold.
A recent test covered by The New Stack compared token consumption when running an identical AI model through three different agent harnesses: Aider, Claude Code, and OpenClaw. The results showed token usage varied by as much as 70-fold depending on the harness, even though the underlying model was the same. This indicates that the agent framework—not the model—plays a dominant role in determining how many tokens are consumed during a task. The New Stack's report highlights that this variation has direct cost implications, since token usage is a primary driver of API expenses for developers using AI coding assistants. The article suggests that developers should consider the efficiency of the harness when budgeting for AI-assisted development, as choosing a more token-efficient tool could lead to substantial savings. The exact benchmark conditions, such as the specific task or model version, were not detailed in the excerpt, but the 70-fold difference underscores the importance of harness selection in managing operational costs.
Harness choice can drastically affect token costs, impacting developer budgets for AI coding tools.