GitHub's usage page says our repo used 554,030 CI minutes in the last 90 days.
That number is wrong, and why it's wrong is most of this story.
I run a small job site, mostly built by AI coding agents that open 30 to 45 pull requests a day. Every one of them runs a full build, unit tests and Playwright e2e. In July I moved all of that off three always-on VMs and onto Cloudflare Containers: one container per CI job, which starts, runs the job, exits and stops billing. Nothing runs while idle.
Here's what four months of that cost, what broke, and whether I'd tell you to do it. The numbers behind every claim are linked at the end, and I've open-sourced the runner.
The numbers
Cloudflare container spend, July 10 to November 10:
- About $19.88/day across the whole period (best estimate; measured $2,108 over 98 of 124 days).
- $34.91/day at the peak, Oct 1-16, when agent PR volume went from 2,555 to 3,125 workflow runs a day.
- $17.77/day after I put a dollar gate on it.
- 97 to 98% of that is the CI runner itself (4 vCPU, 12 GiB).
The three VMs it replaced cost about $300/month and did about 2,536 job-minutes a day. The containers did about 7,312 job-minutes a day in Sep-Oct. That's 2.9x the work for roughly 2x the money.
Then the comparison everybody asks for. For the same 90 days of work, at list prices I checked on November 12:
| Where | Cost |
|---|---|
| Cloudflare Containers (what we were billed, list price) | about $1,608 |
| GitHub-hosted, 4-core for heavy jobs | $6,502 |
| GitHub-hosted, standard 2-core only | $3,288 |
| Blacksmith | $4,322 |
Why so cheap: Cloudflare bills vCPU on what you actually use, and a CI container is idle a lot of its life (waiting on network, on installs, on a browser). Our heavy containers averaged about 41% CPU busy, which works out to $0.0039 per container-minute against $0.012 for GitHub's 4-core runner.
So containers win on the bill, and that is the boring half of the story.
Did it make CI faster?
For a typical PR, no. For the slowest PRs, yes.
I can only compare against the old VMs. We never ran this CI on GitHub's hosted runners, so I don't have that benchmark, and I'm not going to guess one.
ci.yml on pull requests, successful first attempts, start to finish:
| Period | Runs | Median | p90 (slowest tenth) | Median wait before a job starts |
|---|---|---|---|---|
| Old VMs, Jun 26 - Jul 9 | 477 | 17.4 min | 47.6 min | 0.10 min |
| Containers, Jul 10 - 31 | 972 | 19.8 min | 59.4 min | 0.42 min |
| Containers, Aug | 1,117 | 23.1 min | 36.5 min | 0.42 min |
| Containers, Sep | 1,126 | 25.0 min | 39.0 min | 0.18 min |
| Containers, Oct 1 - 17 | 867 | 19.8 min | 29.2 min | 0.18 min |
The median got a little slower. That's not a clean race: the test suite grew over these months, and by September the containers were doing 2.9x the daily work of the VMs. The slow tail is where it changed. On three VMs, a busy hour meant PRs waiting for a free box, and p90 was 47.6 minutes. With one container per job nothing waits for a box, and by October p90 was 29.2 minutes. The first month was the worst of all (p90 59.4), which was the month I was killing my own builds.
The raw rows look like this, one per run:
created_at_utc,era,wall_minutes
2026-07-07T04:47:23Z,Old VMs 06-26..07-09,66.57
2026-07-13T09:00:37Z,Containers 07-10..07-31,18.73
2026-09-05T00:04:05Z,Containers Sep,31.65
2026-10-10T02:27:59Z,Containers Oct 1-17,14.02
All 4,559 of them are in the repo, with the script that makes this table.
What it actually cost
433 commits to the CI system in 125 days. 54% of them were fixes. Something CI-related changed on 85 of those 125 days.
Cloudflare's startup credits paid the container bill, so this barely cost us cash. It cost us engineering.
The cloud bill was never the real cost. My time was. GitHub's own performance page shows some of it: a 6% job failure rate and 51,888 minutes spent on failed jobs, before we even count the work of keeping the runners alive.
Here are the incidents that taught me the most.
1. "The runner lost communication with the server" (July 17-20). About 10 to 14 failed runs a day, no logs. It looked exactly like a Cloudflare flake. It wasn't. GitHub doesn't bind a just-in-time runner to the job it was created for, so a container started for job A would often pick up job B. When job A finished, my code tore down "job A's container", which was by then halfway through job B. I was killing my own builds.
Rule: key every teardown by the job a container is actually running, never by the job that started it.
2. The budget cap that stopped all CI (July 21-22). I had a hard daily minute cap. A bookkeeping bug made the ledger read 9,002 minutes for a day that was really a few hundred. The cap did its job: it refused every container. CI just sat there.
Rule: never put a hard minute cap on CI. Bound spend with structure instead: at most N containers at once, each with a hard per-job deadline, and per-second billing means nothing runs after that.
3. Google said 403 to our containers (Aug 22-24). Playwright downloads Chrome from Google's storage. From Cloudflare's network, that started returning 403. Every e2e job failed. Fix: mirror Chrome in our own R2 bucket and bake it into the image.
Rule: anything a job downloads at runtime is a dependency you don't control. Bake it or mirror it.
4. GitHub Actions stopped starting our runs (Oct 17). An account issue on GitHub's side, nothing to do with our code. From 12:47 UTC every Actions run failed with startup_failure, 511 of them. Self-hosted runners don't help here: if Actions won't start a run, there's no job to hand your runner. Branch protection went too.
The next day I moved orchestration off Actions completely. The Cloudflare Worker now takes the pull_request webhook itself, starts the containers, and posts commit statuses back to GitHub. GitHub stays our git host and PR page. Actions is out of the loop.
Rule: if your CI logic lives in your own code, the CI vendor is replaceable. Ours had to be, overnight.
5. One environment variable too big (Oct 23). Cloudflare Containers cap a single env var at about 16,000 bytes. A long PR description passed in as an env var stopped containers from starting.
Rule: check env size before you start the container, and fail with a message that names the variable.
6. My spend gate read my own notes, not the bill (Oct 18-31). I added a $12/day gate. It read our internal ledger of lane minutes. Real billed spend went over $12 on 8 of 14 days, up to $37.50.
Rule: gate on what the provider bills, not on your own bookkeeping.
The number GitHub can't see
Back to the 554,030 minutes. Since October 18, CI hasn't run through GitHub Actions at all. In those 24 days we ran 15,492 CI lanes and tested 2,364 commits, 104,016 container-minutes, and none of it shows up on GitHub's usage page. 25 of the 90 days in GitHub's window have zero jobs.
If you're using GitHub's dashboard to decide anything about CI cost, check what it can't see first.
One correction
My own commit message from July 22 says "173 runs queued for about 23 hours". When I pulled the actual run data for this post, it didn't hold up: 88 runs created July 20-22 took more than 6 hours from start to finish, and 57 of those still passed. Bad, but not what I wrote at 2 a.m. Pull the raw data before you trust the incident note, including your own.
Should you do this?
Probably not, and I mean that.
Do it if all three are true:
- You'd pay more than about $1,000/month for hosted runners.
- Someone on the team will own it. It is a system, not a config file.
- You need something a vendor doesn't sell. For us that turned out to be CI that keeps running when GitHub Actions doesn't.
Otherwise, buy a managed runner. At our volume it would have cost about 2.7x the cloud bill, and I'd have back a lot of the time those 433 commits took.
The code
I've open-sourced the core: a Cloudflare Worker that takes GitHub's workflow_job webhook and starts one container per job with a just-in-time runner. GitHub App auth, tokens scoped to one repo, a repo allowlist, and public repos denied by default (fork PRs run untrusted code on your runner). Bring your own Cloudflare account; nothing in it is tied to mine. Every lesson above is in docs/LESSONS.md.
Repo: https://github.com/kelvindesman/cf-container-runners
Numbers behind this post: https://github.com/kelvindesman/cf-container-runners/tree/main/docs/postmortem-2026-11
State of the repo, plainly: it's the core extracted from what ran our CI, and it passes its 104 unit tests and secret scans. I haven't yet run it end to end on a fresh Cloudflare account. If you try it, tell me what broke.
If you run CI on something odd, I'd like to hear what broke for you.



Top comments (0)