Field notes · AI Readiness
How Much Does Vibe Coding Really Cost? The 82-Cent AI Coding Problem
sahilpohare@gmail.com · August 6, 2026
If you've shipped a product using Claude Code, Cursor, or another AI coding agent, you already know the feeling: the first few weeks are magic, then everything slows down. Here's the uncomfortable truth backed by data — most of what you're spending on AI-assisted development right now isn't going toward shipping anything at all.
The 18-cent reality
Entelligence AI analyzed more than 1 million pull requests across 2,444 companies and found that for every $1 spent on AI coding tokens, only $0.18 reaches usable, shipped production code.
- Fixing AI-introduced bugs: $0.44
- Rewriting AI-generated code: $0.27
- Review and merge delays: $0.11
- Reaches shipped production: $0.18
Put in real numbers: a team spending $100,000 on AI tokens is getting about $18,000 of actual product value. The other $82,000 is overhead the tool itself created.
Why this happens
It's not a tooling failure — it's a structural one. Addy Osmani's analysis of 22,000 developers found AI-assisted coding pushes defect rates from a 9% baseline up to 54%. GitClear's analysis of over 620 million lines of code found refactoring activity down 70% and code duplication up 81% compared to 2022. Code is being generated faster than it can be structurally absorbed, so it piles up as duplication, entangled logic, and rework.
The part nobody talks about: perception vs. reality
METR ran a randomized controlled trial with experienced open-source developers and found they were 19% slower when using AI tools on real tasks in codebases they knew — yet those same developers believed they'd been about 20% faster.
- Actual task completion time with AI: +19% slower
- Predicted speedup before starting: +24% faster (expected)
- Perceived speedup after finishing: +20% faster (believed)
- AI suggestion acceptance rate: 44%
That gap matters commercially: teams that believe AI is saving them time rarely budget for the cleanup it's quietly generating.
What to actually measure
Cost-per-line is the wrong lens because it only measures what you spent, not what survived. A better framing is ROI per surviving line — how much of what you paid for is still in the codebase 30 days later, since some studies show up to 40% of AI-generated code gets deleted or rewritten within two weeks. Healthy teams are seeing 2.5x–6x ROI on AI coding tools once rework is properly subtracted; anything under 2x after 90 days is a warning sign.
The fix isn't slower AI. It's better boundaries.
The root cause behind most of this waste is architectural, not the model's fault. Practitioners fixing messy Claude Code output consistently land on the same solution: adopt Domain-Driven Design so infrastructure code is clearly separated from business logic, so a single change doesn't ripple across unrelated systems. That's the difference between AI-generated code that compounds into debt and AI-generated code that scales.
If your team is spending most of its AI budget on cleanup instead of shipping, the problem usually isn't the model — it's the missing structure underneath it. Try the calculator to see roughly where your own codebase lands.