Claude Code Costs 2x Pi. Longer Tasks Don’t Fix it.

Last Updated on September 22, 2026 by Editorial Team Author(s): The Dev Loop Originally published on Towards AI. Two 2026 studies disagree on coding agent cost. The turn-count crossover everyone assumes is not there. The real variable is requests per task. Two careful measurements, four months apart, disagree about which coding harness is cheaper. Neither is wrong. They measured different things and both called the result “cost per task.” Photo by Sasun Bughdaryan on UnsplashThe article argues that the supposed “turn-count crossover” between coding agent harnesses is an artifact of measuring cost per task in a way that conflates two different drivers: (1) tokens per request (set by prompt/preamble size and how stable it is, including prompt caching) and (2) requests per task (set by how a harness structures multi-step work, especially batching tool calls to reduce round trips). The author contrasts a UC Berkeley/Arena “HarnessTax” study (finding Claude Code about ~2x Pi on SWE-bench Lite and ~1.5x on Terminal-Bench 2.0) with a separate “Developers Digest” proxy study (showing token overhead differences and that task-level cost parity in July can come from fewer/fatter requests vs …

≈