Vibe coding just got its first real bill.
Earlier this year GitHub Copilot moved to usage-based billing. Premium requests are gone. Now every plan comes with a monthly allotment of AI Credits, and once you burn through them, you pay for what you use, measured in tokens, priced at a cent per credit. Code completions and next-edit suggestions are still unlimited on paid plans. But chat, agent mode, code review, long-running agent sessions, all the expensive stuff, now draws down a meter.
That's a bigger shift than the pricing page makes it look.
For the last couple of years, AI coding felt like unlimited magic. Fire off a prompt, get a wall of code, don't like it, fire off another. The cost of a bad prompt was zero. The cost of not knowing what you wanted was zero. You could brute-force your way to something that ran, even if you couldn't explain a line of it.
Now every one of those moves has a number attached.
- Blindly generating code costs money.
- Debugging the AI's mistakes costs money.
- A lazy prompt that gets a useless answer costs money.
- Re-rolling the same request five times because you didn't think it through. That costs money too.
None of this hurts if you use AI the way a good engineer uses any tool: to move faster on things you already understand. It hurts if AI is your understanding. And that's the part nobody wants to say out loud.
The people this squeezes hardest aren't the engineers who use Copilot as an accelerator. It's the generate-everything-and-pray crowd. The ones who never learned to read the output, only to accept it. When the meter is running, "just keep prompting until it works" stops being a strategy and starts being a spend.
Real developers still come out ahead here, and the reasons are boring on purpose:
- They know what they're building before they open the chat.
- They can look at generated code and spot bad architecture in seconds.
- They debug by reasoning, not by throwing ten more prompts at the wall.
- They use the agent for the tedious 80% and keep their hands on the 20% that matters.
That's the whole difference. AI didn't make engineering skill obsolete. It made it cheaper to fake for a while. Now the faking has a price.
Here's the honest read: this is good for the craft. When compute was free, the market couldn't tell the difference between someone who understood systems and someone who could out-prompt a model. The credit meter draws that line back in. It rewards people who think before they generate and penalizes people who generate instead of thinking.
The era of free, infinite vibe coding is fading. The era of AI-assisted engineers, people who use these tools to go faster at work they could do slower by hand, is just getting started.
If you're a developer, that's not a threat. That's the best news you've had in a while.
This is roughly how we work at devAge. AI is on all the time, but it speeds up decisions we've already made, not the other way around. If you want software built by people who know what the code is doing, let's talk.