• Sources: primary, discussion
  • Summary: The Strands Agents team compared harnesses only on model and benchmark combinations that every harness ran. On Anthropic models, cost per trial ranged from about $0.74 to $1.57, with Claude Code the most expensive and slightly less accurate than the cheapest. The authors credit context offloading, cache reuse and tool-loop efficiency. Dollar figures are partly estimated from token counts, and the authors build one of the harnesses tested.
  • Why it matters: Agent cost depends on the harness as well as the model, so measure tokens per task when you pick one.