Nobody told me that building a side project with AI assistance would come with its own billing problem.
I mean, I knew I'd be paying for API calls. That was obvious. But I had no idea how fast it adds up when you're deep in a build session, iterating on the same feature ten different ways, and Claude is right there, infinitely patient, never telling you "hey, you've already spent $12 today."
I found out the hard way.
The build that got expensive
I was working on TokenBar — a macOS menu bar app that shows you your LLM token usage in real time. The irony of this story is not lost on me.
I spent about three weeks in heavy development mode. Cursor open all day, Claude API calls flying, sometimes switching between GPT-4 and Claude Sonnet mid-session because one felt faster for a particular task. I was in the zone. Making real progress.
Then I got the billing notification.
Not an $80 bill. Not a $200 bill. Just a quiet "your usage has exceeded your threshold" email that I almost marked as read without opening.
It was $340. For one month. On a side project that had zero revenue.
The worst part? I had no idea where most of it came from. I could see the total. I couldn't see the story behind it.
The real hidden cost isn't the money
Here's what I realized after staring at that bill for a while.
The money was recoverable. I'd make it back eventually (hopefully). What I couldn't recover was the confusion about how I was actually using AI during that build.
Was I being efficient? Was I burning tokens on stupid things — asking the same question five times because I didn't like the first answer? Was I using the right model for the right tasks, or just defaulting to the most expensive one out of habit?
I had no idea. And that meant I couldn't improve.
When I was learning to code, I tracked everything. Time spent, what I was working on, where I got stuck. It made me measurably better. But with AI assistance, I'd completely abandoned that habit. I was just... vibing. And the bill was the receipt.
What I changed
I started tracking my token usage in real time. That's actually what turned into TokenBar — I wanted a way to see the running cost right in my menu bar, like a gas gauge, so I'd never be surprised again.
But more importantly, it changed how I think about AI in a build session.
Now when I catch myself re-prompting the same thing four times, I stop and ask why. Usually it's because I haven't thought through what I actually want. The AI isn't the problem — my vague prompt is. Token cost became a proxy for thought quality.
High burn rate on one feature usually means I'm confused about something fundamental. Low burn rate usually means I'm in flow and know exactly what I'm building.
The thing nobody talks about in the "build with AI" hype
Every article about AI-assisted development focuses on speed. How fast you can ship. How much you can build solo. The productivity multiplier.
Fair enough. It is faster.
But faster means you can also make expensive mistakes faster. You can go down the wrong path for longer before you notice. You can generate a ton of code that solves the wrong problem, very efficiently.
The cognitive load of a big build session is still there. AI doesn't remove it — it redistributes it. You spend less time writing boilerplate and more time making higher-level decisions. And those decisions now have a cost attached to every iteration.
That's worth knowing before you start.
The practical fix
If you're building something with AI assistance right now, do yourself a favor and set a daily limit that requires you to actually notice when you've hit it. Not a limit that shuts you down — just one that makes you pause.
For me, having TokenBar in the menu bar serves the same function. I glance up, I see the number, I ask myself if what I'm doing is worth the tokens. Usually yes. But sometimes it makes me realize I've been going in circles for 45 minutes and should just step away and think.
The hidden cost of building with AI isn't the API bill. It's building without awareness of how you're actually using it. Get that awareness, and the bill starts to make sense.
Top comments (0)