GitHub Actions runner cost on backmesh
150,000 minutes a month of 4-vCPU Linux jobs cost $1,800 on GitHub-hosted runners at list price, and about $191–$365 on backmesh at spot prices on 2026-11-13, depending on the cloud and the VM type a job gets. backmesh’s figure includes its control plane and leaves out disks and VM boots, so it’s a floor. backmesh’s core features are free; your clouds bill you for the VMs.
The inputs
Section titled “The inputs”- GitHub-hosted:
linux_4_core, $0.012 a minute, GitHub’s list price before included minutes (docs.github.com, read 2026-11-05). - backmesh: each job runs on its own 4-vCPU, 16 GiB
backmesh-4VM, which bills from its boot until the job ends. The prices are spot prices in the lowest zone of one region per cloud, from the catalog on 2026-11-13 (refreshed with the 2026-11-13 VM type reorder):
| Cloud | Region | backmesh-4 types and spot $/h |
|---|---|---|
| AWS | us-east-2 |
m8i.xlarge 0.0583, m7a.xlarge 0.07, m7i.xlarge 0.0497, m6a.xlarge 0.0547 |
| Azure | westus2 |
Standard_D4ads_v5 0.0381, Standard_D4as_v5 0.0318, Standard_D4s_v5 0.0355 |
| GCP | us-central1 |
n2-standard-4 0.1165, n2d-standard-4 0.0899 |
- The control plane: two always-on VMs with their disks and public addresses, about $86 a month on AWS, $111 on Azure and $74 on GCP. An idle installation pays only this, since runners have no VM while idle.
How it’s computed
Section titled “How it’s computed”GitHub-hosted: 150,000 minutes × $0.012 = $1,800.
backmesh, for each cloud: the control plane plus 150,000 minutes at the cheapest and at the most expensive type’s spot price, $/h ÷ 60 a minute.
| Cloud | Spot $/h | Per 1,000 minutes | 150,000 minutes | Control plane | Month |
|---|---|---|---|---|---|
| AWS | 0.0497–0.07 | $0.83–$1.17 | $124–$175 | $86 | $210–$261 |
| Azure | 0.0318–0.0381 | $0.53–$0.64 | $80–$95 | $111 | $191–$206 |
| GCP | 0.0899–0.1165 | $1.50–$1.94 | $225–$291 | $74 | $299–$365 |
The headline range runs from the lowest month, Azure’s $191, to the highest,
GCP’s $365 — unchanged by the 2026-11-13 reorder, since neither Azure’s cheapest type
(as_v5) nor GCP’s priciest (n2) moved. For example, Azure’s low end is
$111 + 150,000 × $0.0318 ÷ 60 = $190.50. The landing page’s chart
shows the same figures.
What’s left out
Section titled “What’s left out”- Boots. Each job’s VM bills from its boot: about 30 seconds on AWS and 60 to 70 seconds on GCP and Azure. Per 1,000 jobs, that adds $0.41–$0.58 on AWS, $0.53–$0.74 on Azure and $1.50–$2.27 on GCP.
- Disks. Each runner has a 100 GB disk while its job runs. On AWS it’s gp3 at 4,000 IOPS and 500 MiB/s, about $0.038 an hour: $0.011 for storage and $0.027 for IOPS and throughput above gp3’s baseline (EBS pricing).
- Runners started ahead of predicted or cued jobs that no job takes. Each bills for its boot and a short idle time.
- Buckets. The checkpoint bucket and the cache buckets bill for what they store.
- On-demand retries after a spot reclaim.
Spot prices change by the hour and by zone, and other regions differ. On demand cost about 1.7 to 5.4 times spot on 2026-11-13, depending on the type.
The GPU labels by the same method, at list prices: arithmetic, not measured.
- GitHub-hosted: the Linux GPU runner,
linux_4_core_gpu(4 vCPUs, one Tesla T4), $0.052 a minute (docs.github.com, read 2026-11-18). - backmesh: the catalog’s prices on 2026-11-18, spot in the lowest zone, in the same regions as above. GCP has no GPU labels.
| Cloud | Type | Spot $/h | On-demand $/h | Spot per minute | On demand per minute |
|---|---|---|---|---|---|
| AWS | g4dn.xlarge |
0.2732 | 0.526 | $0.0046 | $0.0088 |
| AWS | g4dn.2xlarge, when no xlarge is free |
0.3696 | 0.752 | $0.0062 | $0.0125 |
| Azure | Standard_NC4as_T4_v3 |
0.1766 | 0.526 | $0.0029 | $0.0088 |
10,000 GPU minutes a month cost $520 on GitHub (10,000 × $0.052), and on backmesh $46 on AWS spot (10,000 × $0.2732 ÷ 60), $29 on Azure spot and $88 on demand on either, at each cloud’s first type. The control plane is the same one the other labels use, so it adds nothing. Left out, as above: disks, on-demand retries, and boots, which are longer than the other labels’: a GPU VM boots a larger image and pulls the runner’s image after it, about 90 seconds more.
Estimate your own
Section titled “Estimate your own”- Count your jobs’ minutes a month at each size.
- Run
backmesh install --org <org> --plan, with--runners <cloud>for each cloud to compare. It prints each of those clouds’backmesh-4spot and on-demand prices and the control plane’s monthly cost, without changing anything. Other sizes scale with their vCPUs:backmesh-8costs about twicebackmesh-4. - Multiply each size’s minutes by its price ÷ 60, add the control plane once, and add a boot per job if you run many short jobs.
Once jobs run, backmesh status shows each org’s month-to-date
runner cost by cloud.