Why Vibe coding is eating software

I’ll admit it: Claude Opus 4.6 and GPT-5.3-codex can probably code faster and better than me.

A lot of my coding today is basically guiding agents toward code structures I already have in my head. I want this React component and that component to talk through shared context. Go. The code that comes back is often cleaner than what I would have coded myself, and it is written in a fraction of the time.

The models are smarter. Tokens are cheaper. And because of that, I use them more. Hello, Jevons Paradox.

Back in 2023, we lived in a world where GPT-4 cost around $34 per million tokens. Today, models like Gemini Flash 2.0 are both smarter and cheaper at roughly $0.175 per million tokens. (Token usage has exploded over 15x since 2024. Context windows have expanded too. GPT-3 started at 2,048 tokens. Now it is normal to see models with one-million-token context windows.

Coding has never been easier.

But coding itself was never really the bottleneck. The bottleneck was communication.

One person has requirements in their head and needs to hand them off to someone else. A company has an existing codebase, and a new engineer has to internalize not just how the system works, but why it was built that way. Someone needs to make a change without breaking invisible assumptions, undocumented edge cases, or weird business logic buried in the middle of the stack.

Software development best practices like design reviews exist for a reason. They slow teams down by getting more things written down, clarified, and shared. They help multiple people work together on the same system. They make handoffs possible between product, design, engineering, and everyone else involved. They create room to catch mistakes early, before the organization commits to the wrong product, the wrong workflow, or the wrong abstraction.

Vibe coding bypasses a lot of that, for the good and bad.

Now, one person can drive the whole software development cycle. Internal tools, one-off dashboards, workflow automations, client-specific integrations, prototypes, back-office apps, reporting layers, wrappers around other services. They can pretty much one-shot them. And they will want more.

The world already produces an absurd amount of code. Modern Stack estimates that we produce around 90 billion lines a year. My bet is most of that is mostly plumbing: reading from APIs, transforming data, displaying information, and sending it somewhere else. Simple CRUD apps. And everyone will want more.

We will probably 10–100x our collective code output in the next few years. Hello, Jevons Paradox again.

Software is becoming more disposable. The barrier to productivity is much lower. Starting a new piece of software is beginning to feel as easy as opening a new spreadsheet.

Coding is a new consumable.

What a time to be alive!