January was easy. I barely used any AI tools. Claude was a curiosity, not a dependency. My Anthropic bill: $0.
By March, I was building two macOS apps as a solo dev. Claude was open in one window, Cursor in another. I barely thought about the cost — I was moving fast, shipping features, feeling productive.
Then I opened my billing page.
$150. In one month.
I sat with that number for a second. It wasn't catastrophic, but it was real. And the thing is, I had no idea how I got there. I couldn't point to a single session and say "that's where it happened." It just... accumulated.
The problem with AI billing is it's invisible
With SaaS tools, you pay a flat rate and forget it. With token-based AI APIs, you're charged by the word — essentially. Every prompt, every response, every long context window. And unlike your AWS bill (which at least has a dashboard), most developers I know just... don't check.
You're in the zone, writing code, getting help. You're not thinking about the meter running.
But it is running.
What actually drove my bill up
After I dug into my usage logs, a few culprits stood out:
Long context sessions. I was doing these marathon debugging sessions where I'd paste an entire file, get a response, paste it again slightly modified, get another response. Every round trip was hundreds of tokens. Multiply that by a dozen sessions a day.
Model creep. Somewhere around February I switched from Sonnet to Opus for a few tasks because the output was better. Opus is roughly 5x more expensive per token. I didn't consciously notice the switch — I just kept using whatever felt right.
The "let me just ask" reflex. This is the big one. When AI is right there and frictionless, you stop trying to figure things out yourself first. Every "hey, how do I do X" is a few cents. Ask 50 of those a day and you're at a couple of dollars before lunch.
I needed to actually see what I was spending
The Anthropic dashboard shows you totals. It doesn't show you what's happening in real time while you're working. You don't see the meter ticking. You just see the invoice at the end.
That's what made me build a little awareness tool for myself — something that showed me my token usage as I worked, not after the fact. I wanted the equivalent of a fuel gauge, not just the gas station bill.
I eventually turned that into a proper menu bar app called TokenBar. It sits in your macOS menu bar and shows you exactly what you're spending as it happens — across Claude, OpenAI, and other providers. $5 one-time, no subscription.
The moment I started seeing the numbers in real time, my behavior changed. Not because I got scared of the cost, but because visibility made me more intentional. I started finishing my own thoughts before asking AI. I started keeping context tighter. I switched models deliberately instead of by default.
My April bill is tracking around $60. Same amount of work, more awareness.
The shift that actually helped
Here's the honest answer: the tool matters less than the habit of looking.
Once I started treating token usage like I treat battery life — something to glance at occasionally, not obsess over — I naturally became a better user of these tools. I stopped wasting context on stuff I could figure out in 30 seconds. I stopped keeping giant sessions open when I wasn't actively using them.
If you're building something with AI assistance, even 5 minutes a week checking your usage is worth it. The bills don't feel real until they are, and by then you've already spent the money.
Anyway. $150 was my wake-up call. Maybe this saves you from needing one.
Top comments (0)