Everyone's vibe coding right now. You prompt, the AI writes, you ship. It's legitimately amazing.
I built an entire macOS menu bar app this way. Thousands of lines of Swift I never would have written myself. Claude handled the hard parts, I handled the product decisions, and it came together faster than anything I'd built before.
Then I got my billing statement.
$200+ in one month. On a side project.
Here's the thing about vibe coding that nobody really talks about: when the AI is doing most of the work, you lose track of how much work the AI is actually doing.
I was iterating fast. Asking Claude to refactor this, fix that, try a different approach. Each prompt was maybe a few cents. But I was prompting constantly — debug sessions that went 30 messages deep, exploratory conversations where I'd describe a feature three different ways before landing on the right one, pasting in whole files for context.
It adds up in a way that feels invisible until it doesn't.
A normal coding day, I might write 200 lines manually. With AI, I was generating 2,000 lines of candidates before keeping 500. That's 10x the token volume, and I had no idea.
The part that got me wasn't even a single expensive session. It was the slow bleed.
There's no heartbeat. No "hey, you've spent $50 this week" notification. You just... run a context-heavy debug loop on a Thursday afternoon and somewhere in that fog you've spent $30 without noticing.
I started thinking about this like the early internet era of RAM. When RAM was expensive, good developers knew their memory usage intimately. They watched it, managed it, made deliberate tradeoffs. Then RAM got cheap and we all stopped caring.
Tokens are expensive again. But we're acting like RAM is free.
After the bill landed, I built a small tool for myself: a menu bar app that shows your live token usage as you work. Current session. Today's total. A rough cost estimate.
Nothing complicated. Just visibility.
It immediately changed how I code with AI. Not by making me use it less — I'm still vibe coding constantly — but by making me more intentional. I started noticing which kinds of prompts were token-heavy and which weren't. I started batching context instead of re-pasting the same files over and over.
The $200 month dropped to $60. Same output, better awareness.
I ended up shipping that tool publicly: TokenBar. It sits in your menu bar and shows you what's happening with your AI usage in real time. Supports Claude, OpenAI, and local models via the API.
Vibe coding is here to stay. The economics of building with AI are genuinely different now, and the tools we use to manage those economics need to catch up.
Keep shipping. Just know what it's costing you.
Top comments (0)