AI-assisted coding is 9.9x cheaper, not 19.4x—study fixes flawed math
Two hidden errors inflated AI dev costs by 2x; corrected ratio reveals the truth.
A new arXiv paper from Victor Barros de Miranda Neves, Kiev Santos da Gama, and Vinicius Cardoso Garcia (accepted at ISE 2026 @ CBSoft 2026) presents early results from an ongoing case study on the true cost of AI-intensive software development. A six-person student team spent one academic term building a full conversational onboarding assistant—featuring RAG-based code chat, guided tours, dependency graphs, and technical-debt analysis—using pervasive AI assistance. The researchers instrumented the process with a three-layer cost model: real AI spend, self-reported human effort, and a human counterfactual (what the work would cost without AI).
the initial analysis reported a striking 19.4x cost ratio favoring AI-assisted development. However, a follow-up audit uncovered two independent errors: inferring per-token costs under a flat-rate subscription, and pricing the human counterfactual with incorrect regional labor rates. Together, these mistakes inflated the ratio by roughly 2x. After corrections, the more realistic figure is approximately 9.9x—still a massive advantage, but nowhere near the original claim. The authors argue these errors are easy to make, invisible in final numbers, and likely common in similar reports, making this correction a generalizable lesson for anyone attempting to benchmark AI development costs.
- Six-person team built an AI-heavy onboarding assistant (RAG chat, tours, dependency graphs) in one academic term
- Initial 19.4x cost advantage was corrected to ~9.9x after fixing per-token cost and regional labor rate errors
- Three-layer cost model separates real AI spend, human effort, and counterfactual human-only pricing
Why It Matters
Flawed AI cost comparisons mislead software budgeting; this study offers a robust methodology and a warning.