09/10/2026
Every Job Bills Rounded Up to a Whole Minute
The bill that didn't match the receipt
I've been running more of my own work through AI coding agents lately -- Claude, GitHub Copilot, and Google's Antigravity, each doing a different job in the same afternoon. It's the kind of setup I'd recommend to any business owner sizing up whether "add some automation" makes sense. Mostly, it does. But I recently learned that a tool's sticker price is never the whole story.
Somewhere in the fine print of CI/CD billing (the automated-check systems that run every time code changes) is a detail almost nobody reads. Per GitHub's own Actions billing documentation, every job bills rounded up to a whole minute. Not per hour, not per second, but per job -- rounded up!
How six quick checks become six minutes
- Six checks runEach one takes about fifteen seconds.
- Real work doneNinety seconds, all together.
- Each job rounds upFifteen seconds bills as one full minute.
- What you pay forSix full minutes.
Each job is rounded up on its own, so many short jobs cost far more than the work they actually did.
The multipliers nobody prices in
Then there are the multipliers, which is where it gets genuinely expensive if you're not paying attention. Per GitHub's Actions runner pricing reference, Linux runs at the standard rate, Windows costs double, and macOS costs ten times as much per minute. A wide test matrix that looks cheap on paper -- "let's just check it on every OS" -- can cost far more than the actual time would ever suggest.
The two multipliers aren't the same kind of markup, either, and knowing why explains a lot about how cloud pricing works in general.
Same minute, three prices
Linux -- the standard rate
The baseline rate, and the one the other two multiply. Linux shares commodity servers cheaply.
Windows -- double
Microsoft charges GitHub a licensing fee on top of otherwise-ordinary server hardware. It's a pass-through cost, nothing more.
macOS -- ten times
Apple's license only allows macOS on genuine Apple hardware, so GitHub has to lease and maintain a real Mac for every runner.
That last one is worth a second look. There's no cheap way to pack dozens of virtual macOS machines onto one commodity server the way Linux and Windows can share one. The 10x isn't GitHub being greedy. It's what real Apple hardware actually costs at scale.
Why this is a business lesson, not a dev lesson
I didn't write this because it's a quirky billing detail. I wrote it because this mistake can happen anywhere you compare tools on their sticker prices instead of on how the meter actually runs. AI subscriptions, cloud hosting, ad platforms -- they all have a version of "rounds up, and multiplies more than you'd guess."
The fix isn't avoiding automation. It's reading the meter and testing before you scale anything up.
That's the whole pitch behind what I do now: let my Claude help your Claude. Not because AI is magic, but because the people who've already been burned by the fine print can save you from being burned by it too. Like me, more than a few times!
If you're weighing whether AI-assisted systems make sense for your business, keep in mind that a competitor is weighing it too. I've been amazed at how fast business development, planning and execution have improved across so many parts of a business.
Small businesses have the most to gain.
Let my Claude help your Claude
If you think AI will need to be part of your business's future growth strategy but you're not sure where to start, I offer affordable walk-throughs and will gladly come to your business if you're here locally in Jacksonville, Florida, or we can hop on a video call. No charge to ask me a few questions for a 30 minute chat. For AI specifically, that's me, personally -- broader business growth and technology consulting is Brandager's own team.