Vibe Coding in 2026: What the Data Shows After 16 Months

TL;DR

• Andrej Karpathy coined "vibe coding" on February 2, 2025. By mid-2026, he calls the term passé and has moved to a new framework.
• A July 2025 METR study found experienced developers are 19% slower with AI tools, despite predicting 24% gains.
• 91.5% of vibe-coded apps carry at least one security vulnerability, per Q1 2026 research across five independent studies.
• Vibe coding works well for: MVPs, internal tools, and non-developers validating ideas fast.
• It reliably fails for: anything touching authentication, payments, or user data without expert review.
• The professional evolution is "agentic engineering" — same tools, but the engineer reads every diff and stays accountable.

On February 2, 2025, Andrej Karpathy fired off what he later called a "shower of thoughts throwaway tweet": "There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."

That tweet became Collins Dictionary's Word of the Year for 2025. The market built around it is now worth 4.7 billion. Lovable, one of the leading tools, reached 200M ARR and a $6.6B valuation by mid-2026. Y Combinator's Winter 2025 batch had 25% of startups running on codebases that were 95% AI-generated.

Sixteen months is long enough to stop asking "will this work?" and start asking "what does the data say?" The answer is messier than the hype, more useful than the backlash, and genuinely different depending on who you are.

The Productivity Numbers Don't Tell the Whole Story

The headline stats are easy to find: 55% faster project completion, 5.8x faster application development, 74% of developers reporting productivity gains. What those summaries consistently omit is the July 2025 METR study.

It tested experienced software engineers on real tasks with and without AI coding tools. The result: they were 19% slower with AI assistance, despite confidently predicting they would be 24% faster. The expectation-to-reality gap was 43 percentage points in the wrong direction.

That is not an argument against AI tools. It is a signal about who actually benefits. The productivity gains are real but highly uneven:

• Senior developers with 3+ years of experience report gains of 40-81% in tasks measured by third parties
• Junior developers show no statistically significant output improvement
• Non-developers, who represent 63% of active vibe coding users (product managers, founders, designers), gain something different entirely: access they did not have before

The METR result makes sense when you think about what AI tools actually do. They reduce friction on well-defined, modular tasks. Experienced engineers spend most of their time on architecture, debugging, and judgment calls, not typing code. Typing code faster does not help when the bottleneck is thinking. Meanwhile, a founder who could not build a prototype at all before can now build one in a weekend.

The tool is genuinely useful. It is not universally useful in the same way for the same people.

The Security Problem Is Structural, Not Incidental

The vulnerability statistics come from multiple independent sources and cluster around the same conclusion. OX Security reports 62% of AI-generated code ships with security flaws. A Q1 2026 analysis published by SoftwareSeni found 91.5% of vibe-coded applications carry at least one vulnerability across five independent studies. Escape.tech scanned 5,600 publicly available vibe-coded apps and documented over 2,000 high-impact vulnerabilities, 400-plus exposed secrets, and 175 instances of exposed personal data in live production systems.

The 91.5% figure sounds alarming. It is. But understanding why is more useful than just being alarmed.

AI generates code based on what you prompt for. When you write "build me a todo app with user accounts," the AI writes auth that functions. It will not add SQL injection protection, rate limiting on login attempts, or CSRF tokens unless you specifically ask. Those security layers require domain knowledge to know they exist.

Escape.tech named this pattern directly: vibe coding failures tend to be structural, meaning entire security layers that were never implemented because the AI was never prompted to implement them. This is different from the AI writing broken code. The AI wrote exactly what was asked. The missing pieces were never requested.

This is fixable by people who know what to ask. It is not fixable by hoping the AI will handle it unprompted.

CodeRabbit's December 2025 analysis found AI co-authored code contains approximately 1.7 times more "major" issues compared to human-written code, including a 2.74 times higher rate of security vulnerabilities. That data covers all AI-assisted code, not just pure vibe coding apps, which suggests the problem scales with how much AI oversight you remove.

Understanding the Tool Spectrum

"Vibe coding" has become shorthand for any prompt-driven development, which creates confusion because the tools serve very different users.

No-code AI builders (Lovable, Bolt.new, Replit Agent) let you describe what you want in plain English. The platform generates, hosts, and manages the code. You never see the underlying files. These are designed for non-developers and for rapid idea validation.

AI code editors (Cursor, Windsurf, Claude Code) require programming knowledge. They generate code inside your development environment, where you read and review every change before committing. The AI is a very fast, sometimes unreliable collaborator, not an autonomous builder.

The failure mode most teams hit: starting with Lovable, reaching 80% of a working MVP, then trying to extend it into a production system without understanding what was built underneath. No-code-generated codebases often have patterns optimized for the generator, not for long-term maintainability.

A practical decision framework:

Karpathy Has Already Moved On

The person who coined the term now says it is passé.

At Sequoia Ascent 2026, Karpathy introduced a framework he calls "agentic engineering." His argument: by December 2025, model reliability improved enough that the real bottleneck shifted. The question is no longer whether AI can write code. It is whether humans can maintain meaningful oversight of AI-written systems.

The distinction he draws:

Vibe coding raises the floor. Anyone can build something functional. Code quality is often what Karpathy calls "bloaty, lots of copy-paste, awkward abstractions that are brittle." It ships, and that's valuable. Low accountability is appropriate because stakes are low.

Agentic engineering raises the ceiling. Professional developers orchestrate AI agents while maintaining full accountability. They design specs before prompting, review diffs before merging, write tests, build eval loops, and manage permissions carefully.

The most-quoted line from the talk: "You can outsource your thinking but you can't outsource your understanding."

Practically, the line between the two approaches is not which tool you use. It is what you do after the AI generates code. A developer using Cursor who accepts all suggestions without reading them is vibe coding. A developer who treats every generated diff like a PR from a junior engineer, reads it line by line, and refuses to merge anything they cannot explain, is doing agentic engineering. Same tool. Different discipline.

What to Always Write or Review Yourself

Based on the documented failure patterns, some categories should never go to production without human review regardless of approach:

• Authentication logic: Password hashing, session management, token validation
• Payment processing: Webhook handling, idempotency keys, error state recovery
• Database queries with user input: Any code that constructs queries from external data
• Authorization checks: Who can read, write, or delete which records
• Secret handling: Environment variables, API key rotation, token storage

Everything else is a reasonable candidate for AI-assisted generation with review. These five categories are where the structural failures cluster.

FAQ

Can non-developers actually ship production apps using vibe coding?

Yes, with an important distinction about what "production" means. An internal tool used by 5 people is different from a consumer app storing payment data for 5,000. Non-developers are building and shipping the former category successfully. For the latter, the consensus from security researchers in 2026 is: validate your idea with vibe coding, then have a developer audit the security-critical paths before you serve real users with real data.

What is the actual difference between Lovable and Cursor?

Lovable abstracts away the code entirely. You describe, it builds, you ship, you never touch the files. Cursor is a code editor where AI helps you write faster, but you remain in the codebase, reading and committing every change. The right choice depends on whether you can read code, not on which tool has better reviews. For developers, Cursor. For everyone else starting out, Lovable or Bolt.

Is vibe coding dead?

As a practice, no, it is mainstream. GitHub reports 46% of new code is now AI-generated. As a term that accurately describes what experienced engineers are doing with these tools, it is fading. Karpathy himself moved on from it. The tools are the same; the professional expectation of oversight has changed.

What does vibe coding actually cost to start?

Lovable's paid plan starts at 25/month (as of June 2026). Cursor Pro is 20/month. Bolt.new has a free tier with limited tokens per day. Claude Code is API-usage-based, priced through Anthropic's API. A solo founder building an MVP can typically stay under 50/month at low usage. Production workloads on large codebases can run 200-500/month depending on how heavily you use token-intensive operations.

The Actual Verdict

Vibe coding is not a fraud and it is not a replacement for software engineering. It is a spectrum: on one end, non-technical people building real things for the first time; on the other, experienced engineers using AI as a high-velocity collaborator while maintaining accountability for what ships.

The failures are predictable. They follow a pattern: accepting large amounts of generated code without reading it, skipping security layers that were never prompted, and trying to scale a no-code prototype into a production system without understanding the underlying architecture.

The people getting the most from these tools in mid-2026 are not the ones who "fully give in to the vibes." They are the ones who understand exactly when to delegate to AI and exactly what to inspect when it comes back.

For an overview of the broader AI coding tool landscape, our best AI coding assistants guide covers the full spectrum from Cursor to GitHub Copilot with pricing as of 2026.

If you are starting to build agent-driven workflows beyond just code generation, how to use AI agents walks through the practical orchestration patterns that separate one-off generation from repeatable automated systems.