GitHub Actions runners, for a fraction of the price.
FlameWill runs each job on a fresh ARM machine sized by the memory it needs. Up to 8× cheaper than Blacksmith and up to 16× cheaper than GitHub-hosted runners. Switch with one line.
- runs-on: ubuntu-latest + runs-on: flamewill-arm-2c8g
- Set up job1s
- Run actions/checkout@v56s
- Set up node, pnpm, browsers1m 39s
- Build everything, then the fast gate3m 30s
- Post setup13s
- Complete job1s
Per minute, against the cheapest runner on each that holds the same job’s memory; prices read on 24 Oct 2026. FlameWill’s are launch prices: it is in preview and bills nothing yet.
Real CI, already running on FlameWill
Our first customers moved their workflows in October 2026. Same jobs, one label changed.
Build, Docker checks, API tests and deploy. 33–35 min on GitHub-hosted runners.
6 to 18 min on GitHub-hosted runners.
1.3 GB of node_modules restored from the cache instead of installed.
Cheap where it can be. Fast where it counts.
The savings come from what you stop paying for: idle cores, idle machines, paid disks. Not from slower jobs.
Sized by memory, not by cores
Most CI jobs use one or two cores and run out of memory first. Pick the memory your job needs and stop paying for cores it never uses.
One line to switch
Same workflows, secrets, environments and OIDC. To go back, change the line again.
jobs: test: runs-on: flamewill-arm-2c16g
Caches next to the machines
Your npm cache comes back before the first step and is saved after the last, with no setup. Keep anything else, like node_modules, browsers or toolchains, with one command.
fw-cache restore "$key" node_modules # hit: 1.3 GB in 6 s
The cost of every job, in its log
Every job ends with its peak memory, CPU time and what its machine cost, so you can size each job by evidence and spot the expensive ones.
peak process memory 1.81 GB of 7.6 GB machine cost ≈ $0.0022
Fast local disks
Your job works on local NVMe, up to 220 GB, included in the price.
A fresh machine every time
One job per machine, destroyed after it. No inbound network, no SSH keys. Machines run in the EU.
Ready for real work
Already installed, so your setup steps download less:
FlameWill, GitHub-hosted and Blacksmith
Price per minute of the cheapest runner that can hold the job, and what each one does.
| Compared | FlameWill | GitHub-hosted | Blacksmith |
|---|---|---|---|
| Price for a 2.5 GB job | $0.0006/min | $0.005/min | $0.0025/min |
| Price for a 13 GB job | $0.001/min | $0.012/min | $0.008/min |
| Price for a 15 GB job | $0.0013/min | $0.022/min | $0.010/min |
| Memory per vCPU | up to 8 GB | 4 GB | 3 to 4 GB |
| Linux on ARM64 | |||
| Linux on x64 | Coming | ||
| Windows and macOS | |||
| Dependency cache with no workflow change | Faster actions/cache | ||
| Docker layer cache | Coming | ||
| Cost and peak memory of each job | In the job’s log | In a dashboard |
Prices from each vendor’s own price page on 24 Oct 2026: GitHub, Blacksmith. GitHub does not state its larger runners’ memory; we assume the 4 GB per vCPU of its standard ones. FlameWill’s prices are launch prices.
Three sizes, priced per minute
Billed per second of the job’s machine. Nothing runs, or costs, while you have no jobs. Launch prices, not billed yet
Questions
Do I need to change my workflows?
Only the runs-on line of each job you move. Secrets, environments, OIDC and actions work as on GitHub’s runners, because they belong to GitHub, not to the machine.
Is it as fast as Blacksmith?
Not on every job. Blacksmith boots machines in seconds; ours take about 45 to 60 seconds, because we keep no idle machines running at your expense. Once running, our customers’ jobs have run as fast as or faster than on GitHub-hosted runners. If speed on one core matters most, move your slowest job and compare.
Why ARM?
AWS Graviton gives the most memory and compute per dollar. Node, Python, Go, Rust, Java and Docker images with an arm64 build run natively. An x64-only binary or image does not, yet.
What if AWS takes a machine back?
FlameWill runs on AWS’s spare capacity, which is why it is cheap. Taking a machine back mid-job is rare; when it happens, the job fails with “runner lost” and you re-run it. A job that cannot get a machine at once waits and is retried automatically.
Where do my jobs run, and who can see them?
On machines in AWS Stockholm, one job per machine, destroyed after it, with no inbound network. Your caches are readable only by your organization’s runners. Your code, logs and artifacts stay where GitHub keeps them.
How do I start, and what does it cost today?
Request access and tell us which repository to start with. FlameWill is in preview: we add organizations one by one, and nothing is billed yet.
Move your most expensive job first.
Change one label, read what the job cost at the end of its log, and decide.