6 min readBy JJ Eaton

Why your CI bill is about job starts, not minutes

Our app ran about 9.7k CI workflows in October. Compute would cost around $9 on one host, but plan limits on job starts would push it past $500.

Card reading 9.7k CI runs in October. Cheap minutes, expensive starts. $9 in compute, $500+ in job starts, over a dark desk with a monitor and a PC

If your team builds software with AI coding agents, your CI bill is about to become a real line item. CI (continuous integration) is the set of automated checks that run every time someone proposes a code change: build it, test it, lint it, make sure nothing broke. Everyone is shopping for the cheapest CI host right now, and most people compare one number: the price per minute.

After two months of fighting our own CI costs, I think that is the wrong number. For us, the expensive part is not how long jobs run. It is how many times a job starts each day.

What our usage actually looks like

Lockpick, our app, is relatively small. In October it ran about 9,700 CI workflows. Across the month that came to around 248,000 minutes.

The shape of those jobs matters more than the total:

  • The median job took 31 seconds.
  • 61% of jobs finished in under a minute.
  • A third of the billed minutes were just rounding, because a 31-second job gets billed as a full minute.

That pattern comes straight from how we work now. I contribute anywhere from 30 to 80 pull requests a day, working with AI agents. Every one of those triggers checks. Even with heavy caching, TurboRepo on our web app, and stacking draft pull requests to cut down runs, the number of job starts keeps growing.

The pricing trap

On paper, the cheapest host I looked at was boat.dev, at roughly $0.0003 per minute. Based on our usage, the compute itself would cost around $9 a month.

But plans cap how many jobs can start per day. On my number of starts, the extra starts on the $20 plan would come to more than $500. Even the $200 plan wouldn't cover our usage.

Short jobs plus lots of starts means you pay for the plan, not the compute.

When I ran the rest of the numbers, strictly minutes to minutes and based on last month's usage, it looked like this:

OptionEstimated monthly cost
GitHub Actionsabout $630
EC2 runners with GitHub (what we use now)$342
boat.devabout $500, limited by starts, not compute
Vercel CIabout $600 to $930

The Vercel figure also doesn't include per-seat pricing or any other add-ons we would need to configure. None of these are huge numbers on their own, but this is one small app, and the trend is only going up as agents write more code.

The fix that worked best: a faster language

The single biggest reduction in our CI and hosting costs has not come from switching providers. It has come from moving services away from Next.js and TypeScript.

When we moved one service from TypeScript to Go, its CI time dropped from more than five minutes to around 30 seconds, and its memory use dropped about 10x. That is the kind of saving no pricing page will give you.

I have spent around 10 hours every week for the last two months trying to optimise CI usage. Realistically, I could have reorganised things into fewer, longer-running jobs. But the bigger lesson is that compute is not the only part of the problem. The language you build in and the way your workflows are designed matter just as much.

The three ways out

As I see it, there are three real options:

  1. CI providers raise their job-start limits. I think they need to increase start counts on all plans by three or four times, or expect most customers who work like this to leave after the first month.
  2. Run CI on your own hardware. Buy or rent machines and run the workflows yourself. This is what I built Davit for: basically Boat, but on your own hardware.
  3. Turn off remote CI and have agents run checks locally. If the agents writing the code also run the tests on your machines before anything is merged, you don't need to pay someone else for every start.

I am doing a combination of two and three. My agents run all of our staging CI workflows locally on my own hardware. Some remote CI still runs for production, on a 9950X machine I control. I am even starting to ask whether we need traditional CI at all.

What this means for your business

If your developers are starting to use AI coding agents, or you are paying an agency that does, here is what I would take from this:

  • Look at job starts, not just minutes. Ask your team or your provider how many CI jobs start per day and how long the median one takes. If most finish in under a minute, per-minute pricing is hiding the real cost.
  • Expect volume to jump with AI. Agents open far more pull requests than people do. A plan that fits a five-person team today may not fit the same team with agents next month.
  • Watch the rounding. If a third of your minutes are rounding, you are paying for time nothing is running.
  • Consider your own hardware. A single capable desktop machine can run a lot of short jobs with no per-start limits, and your code never leaves your control.
  • Treat language and architecture as cost decisions. Faster builds are cheaper builds. Moving a busy service to a faster language can cut CI time from minutes to seconds.
  • Make the agents do the checking. If an agent wrote the change, it can run the tests locally before a person ever reviews it.

The point isn't that any one provider is bad. It's that pricing built for humans opening a handful of changes a day doesn't fit teams where AI opens dozens.

Where to start

If your software spend is creeping up as your team adopts AI coding tools, it's worth a proper look before you switch providers or sign a bigger plan. That is part of what we cover in a Build Things AI readiness audit: how your team actually builds and ships, where the money goes, and whether running things on your own hardware makes sense for you.

Sources