The $800 Lesson in Vibe Coding

Matthew Berman, an AI coding YouTuber, received an $800 Vercel bill after two weeks of deploying AI-generated code. He hadn’t reviewed the deployment configurations—his AI assistant had recommended Vercel and chosen the defaults, including the most expensive build tier and concurrent builds enabled.

“Very similar to me not reading any of the code, I didn’t think much about the services I was using either, how they were set up, or any of the configurations,” Berman said. “I just said deploy.”

The incident is a clean example of what’s now called “vibe coding”—the practice of describing what you want in natural language and letting AI agents write, deploy, and iterate without line-by-line human review. The approach has accelerated shipping speeds dramatically. It has also introduced a category of risk that didn’t exist when humans typed every character: infrastructure costs chosen by AI agents optimizing for speed, not spend.

The Vercel problem

Vercel’s billing is usage-based, charging for data transferred, requests made, and compute duration. Berman’s AI had selected the Turbo build tier and left concurrent builds enabled, meaning every deployment ran parallel builds that multiplied his costs. After manually optimizing—switching to sequential builds, using GitHub hooks, dropping to a lower tier—his builds went from three to four minutes down to ~1 minute, and costs dropped to “a couple dollars.”

The fix wasn’t complex. The problem was that Berman never knew to look. His AI assistant had recommended Vercel, configured it, and deployed to it. He trusted the agent’s judgment the same way he trusted its code.

The IDE shift

Berman’s experience arrives as the major AI coding tools redesign their interfaces to de-emphasize code review entirely.

Cursor 3, released April 2, 2026, introduced an “Agents Window” that the company describes as “simpler, more powerful, and centered around agents.” The new interface lets users run many agents in parallel across repositories and environments. Cloud agents were removed from the editor and moved to this separate window. A “Design Mode” lets users annotate UI elements directly in a browser view rather than looking at code.

Claude Code, Anthropic’s agentic coding tool, is now available across terminal, IDE, desktop app, and browser. Its documentation emphasizes the natural language interface: “Describe what you want in plain language. Claude Code plans the approach, writes the code across multiple files, and verifies it works.”

The direction is consistent. Chat and visual feedback are primary; code is secondary. The tools are being built for a workflow where humans don’t read every line.

“It is physically impossible to review AI code.”

The review problem

Berman’s framing is blunt: “We are shipping more code than ever… but at the same time, we’re actually understanding the code less than ever.”

He argues that not reviewing code isn’t a bug in these tools—it’s a feature. The interfaces are designed around the assumption that AI will write, test, and verify. Human attention is meant to be spent on product decisions, not syntax.

This is a defensible position when the code works. It becomes a problem when the code makes expensive choices humans didn’t authorize. Berman’s $800 was recoverable. A production database misconfiguration might not be.

The deeper concern Berman raises is speculative but worth noting: if AI writes code optimized for AI to read and extend, there’s a path toward codebases humans cannot understand. We aren’t there yet. The question is whether the incentives—speed, shipping, iteration—push us in that direction.

What to watch

AI coding agents consistently recommend the same services: Vercel, Resend, Fly.io, Railway. Resend recently announced it crossed two million users, up from one million just four months prior. These platforms are benefiting from AI-driven adoption at a rate that suggests many developers, like Berman, are clicking “deploy” without auditing the bill. Whether that changes—or whether surprise invoices become the new cost of shipping fast—is the question for the next six months.