Two developers. Same model. Same task. One pays 5x more and gets the same result. The difference isn't the model — it's the harness. A new study called HarnessTax tested 21 model-harness combinations and found that the software layer surrounding an AI model can move inference cost by as much as 5x without producing a comparable change in task success. The harness is the layer that handles instructions, tools, memory, and context management. When you choose between Claude Code, Codex CLI, or another coding agent, you're choosing a model and a harness at the same time. Three findings worth internalizing: The harness can move your bill far more than your score. Claude Fable 5 on SWE-bench solved 97.8% of tasks with Claude Code and 96.7% with a simpler harness called Pi. The success difference was 1.1 percentage points. The cost difference was 2x. Claude Code's initial context averaged more than 10 times Pi's, driven by longer instructions and larger tool definitions — all of which the model has to process before it even starts your task. A simple harness can compete with feature-rich ones. Pi gives the agent four tools: read, write, edit, and bash. It still hit the cost-success Pareto frontier on both benchmarks — meaning the highest accuracy for a given budget. Simpler isn't always better, but it's a useful starting point before adding scaffolding that may not move your actual results. Your vendor's harness isn't automatically your best harness. In 9 of 12 model-benchmark comparisons, an alternative harness outperformed the provider's own default. GPT-5.6 Sol scored 83.3% on Terminal-Bench with Pi at $0.42 per attempt — compared with 78.9% at $0.76 with OpenAI's own Codex CLI. The practical takeaway: evaluate "model × harness × workload" as one system. Run your own representative tasks, test several combinations, and track success, cost, and latency together. The right harness for high-volume routine work is not the same as the right harness for a task where failure is expensive. https://alphasignal.ai/newsletter/ai-harness-tax-coding-agents-study