FrontierHarness Eval Shows 17x Cost Variance Across Coding Agents
September 2, 2026
FrontierHarness evaluates nine coding agents using the same model, revealing performance scores ranging from 66.7% for Codex to 50.0% for OpenCode. Testing highlights a 17x cost-per-pass disparity among agents despite identical underlying model usage.
HOW THIS AFFECTS YOU
●
builderYou can optimize agentic workflows by selecting agents based on cost-efficiency rather than just raw performance.
●
founderThis shows significant margin opportunities in optimizing agent orchestration costs.