Guide · prices checked 7 November 2026
What a self-hosted GitHub Actions runner really costs
Self-hosting looks cheapest on a price sheet: an AWS spot machine works out to about $0.00075 a minute at 2 vCPU. The bill you actually pay adds boot time, disks, public addresses, idle hours and the time someone spends keeping it running.
The raw machine price, per minute
| Machine | 2 vCPU | 4 vCPU | 8 vCPU | Billing |
|---|---|---|---|---|
| AWS EC2 spot (m7i) | $0.00075 | $0.0015 | $0.0030 | Per second, 60 s minimum, spot price moves hourly; can be interrupted |
| AWS EC2 on-demand (m7i) | $0.0017 | $0.0034 | $0.0067 | Per second, 60 s minimum |
| Hetzner Cloud CCX (dedicated vCPU) | $0.0012 | $0.0023 | $0.0037 | Per started hour |
AWS: m7i.large, xlarge and 2xlarge in us-east-1, Linux, from the AWS price list and a spot snapshot on 7 November 2026 ($0.1008, $0.2016 and $0.4032 an hour on-demand; $0.0449, $0.0897 and $0.1774 spot). Hetzner: CCX13, CCX23 and CCX33 ($0.0697, $0.1386 and $0.2227 an hour). Hourly prices divided by 60.
What gets added to that number
- Boot time is billed. On AWS the machine is billed from start, so the minute or two it takes to boot and register a runner is paid for on every ephemeral job.
- Disk and public addresses. EBS volumes and public IPv4 addresses are billed separately on AWS.
- Started hours. Hetzner bills each started hour. One server per job means a 10-minute job pays $0.0697 at 2 vCPU, about $0.007 a minute, more than most hosted runners.
- Idle time. Keeping warm machines ready so jobs start fast means paying for hours no job runs. A pool that is busy a third of the time costs three times its per-minute price per minute of actual work.
- Interruptions. Spot machines can be reclaimed mid-job; the job reruns and pays again.
- People. Images, autoscaling (ARC or similar), security updates and debugging are engineering time, usually the largest cost and the one that never shows on the cloud bill.
Against hosted runners
The lowest hosted rate is $0.001 a minute at 2 vCPU (Sengi Basic), with boot and teardown unbilled and nothing to operate. That is above a perfectly used spot machine ($0.00075) and below on-demand AWS ($0.0017) and a Hetzner server used for one job per hour. Add any idle time or boot overhead and the spot advantage disappears: at 70% use, the spot machine's effective rate is already about $0.0011 a minute.
| Option | 2 vCPU per minute | You operate it? |
|---|---|---|
| AWS spot, fully used | $0.00075 + disk, IPv4, boot | Yes |
| Sengi Basic (hosted) | $0.001 | No |
| Ubicloud (hosted) | $0.00125 | No |
| AWS on-demand, fully used | $0.0017 + disk, IPv4, boot | Yes |
| GitHub-hosted | $0.006 | No |
Self-hosting still makes sense when jobs need access to a private network, special hardware, or a very large steady load that keeps machines busy around the clock. The software to run it is in the catalog, with licences.
Questions
- Is a self-hosted GitHub Actions runner cheaper?
- Only when machines stay busy. A fully used AWS spot m7i.large works out to about $0.00075 a minute at 2 vCPU before disk, public IPv4 and boot time; idle time, interruptions and upkeep usually push the real rate above hosted runners that start at $0.001 a minute.
- How much does an AWS runner cost per minute?
- m7i.large in us-east-1 is $0.1008 an hour on-demand (about $0.0017 a minute) and about $0.0449 an hour spot ($0.00075 a minute) as of 7 November 2026, plus EBS storage and public IPv4, and boot time is billed.
- Why is Hetzner expensive for CI jobs?
- Hetzner bills each started hour. A short job on its own server pays the full hour: $0.0697 for a CCX13, so a 10-minute job costs about $0.007 a minute.