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.”
The 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 more/thinner ones). They explain that system prompt content is not a one-time intercept because the API is stateless, so it behaves more like a per-request (per-turn) component whose impact is heavily discounted or penalized by cache read/write behavior and whether the harness churns its request prefix. To test their prediction that no crossover should occur for Pi because its per-request payload stays cheaper, they describe building a logging “relay” in front of a model endpoint that records usage from SSE events, and propose a measurement plan where requests per task is controlled; overall they conclude that benchmark turn length alone won’t answer cost questions—parallelism and batching structure within real repositories matter.
Read the full blog for free on Medium.
Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.
Published via Towards AI
Towards AI Academy
We Build Enterprise-Grade AI. We'll Teach You to Master It Too.
15 engineers. 100,000+ students. Towards AI Academy teaches what actually survives production.
Start free — no commitment:
→ 6-Day Agentic AI Engineering Email Guide — one practical lesson per day
→ Agents Architecture Cheatsheet — 3 years of architecture decisions in 6 pages
Our courses:
→ AI Engineering Certification — 90+ lessons from project selection to deployed product. The most comprehensive practical LLM course out there.
→ Agent Engineering Course — Hands on with production agent architectures, memory, routing, and eval frameworks — built from real enterprise engagements.
→ AI for Work — Understand, evaluate, and apply AI for complex work tasks.
Note: Article content contains the views of the contributing authors and not Towards AI.
Related posts
Recent Posts
