Summary: StarSling sells drop-in GitHub Actions runners from $0.004 to $0.128 per minute, billed by the second on new paid plans, plus NVIDIA GPU runners from $0.02522 per minute. Its headline "33% cheaper" compares the 4 vCPU runner ($0.008) with GitHub's 4-core larger runner ($0.012), but a private repo on plain ubuntu-latest pays GitHub $0.006 for 2 cores, so the right comparison depends on what you run today. We checked every rate on the official pages on 19 November 2026 and priced a sample month of agent-heavy CI.
Why CI minutes suddenly matter for agent teams
When coding agents open pull requests, every push triggers your workflows. A team that used to run a few hundred CI jobs a week can end up running thousands of short checks a day: lint, type-check, unit tests, and re-runs after each agent fix. Two things then drive the bill much more than they used to: the per-minute rate of the runner, and how each provider rounds short jobs.
StarSling is a Y Combinator company that positions itself as "agent-native CI for GitHub Actions". It does not replace GitHub Actions. You keep your workflow files and change the runs-on label, for example from ubuntu-latest to starsling-ubuntu-26.04. On top of the runners it offers optimization agents that open pull requests against your workflows (included on paid plans, but off by default for new accounts) and a set of free open-source skills such as /ci-speedup and /ci-score that you can run with your own coding agent.
StarSling CPU runner rates
All CPU runners are Ubuntu 26.04 or 24.04 on 5th Gen AMD EPYC. The pricing page lists per-second prices and normalized per-minute equivalents:
- 2 vCPU / 8 GB (
starsling-ubuntu-26.04-2): $0.004/min ($0.00006667/s) - 4 vCPU / 16 GB (
starsling-ubuntu-26.04, the default): $0.008/min - 8 vCPU / 32 GB: $0.016/min
- 16 vCPU / 64 GB: $0.032/min
- 32 vCPU / 128 GB: $0.064/min
- 64 vCPU / 256 GB: $0.128/min
The pattern is simple: every size is exactly $0.002 per vCPU per minute. GitHub's x64 larger runners are not linear in the same way, which is why StarSling's advertised saving shrinks from 33% at 2 and 4 vCPU to 21% at 64 vCPU.
How StarSling bills time
- New paid subscriptions are billed in whole seconds from job start to finish, with no one-minute minimum. A 5-second job is billed as 5 seconds.
- Existing paid subscriptions stay on whole-minute billing with a one-minute minimum until they are migrated.
- Queue time is never billed, concurrency is listed as unlimited, and there are no seat fees.
- Failed jobs are billed for the time they actually ran.
- Free start: new accounts get 2,000 free minutes in the first month, but they are tracked as 2-vCPU-equivalent minutes, worth $8 of compute. On the 4 vCPU default that is really 1,000 minutes. Unused minutes do not roll over, and no credit card is required to start.
- Storage (StarSling Disks) is separate at $0.50 per GB-month, and the free trial does not include storage.
StarSling GPU runners
Each GPU rate includes the card plus a 4 vCPU, 16 GB RAM, 100 GB disk host:
- NVIDIA RTX PRO 6000 (default GPU label): $0.06422/min
- NVIDIA RTX 5090: $0.03022/min
- NVIDIA RTX 4090: $0.02522/min
- NVIDIA H100: listed as coming soon
GPU access is not self-serve in the same way as CPU: the page asks you to contact the team with the GPU you want and what you plan to run, and they open access for your organization. You pick a card with a selector in the label, for example starsling-ubuntu-26.04-gpu/gpus=rtx-4090:1.
One thing to read carefully: the "92% to 96% you save" figures are not per-minute comparisons. StarSling compares against GitHub's only GPU runner, a 16 GB NVIDIA T4 at $0.052/min, and assumes the same job needs as many more minutes on the T4 as its lower FP32 TFLOPS suggests (14.8x for the RTX PRO 6000). Per minute, the default RTX PRO 6000 is actually more expensive than GitHub's T4. Whether your job really finishes 10x to 15x faster depends on whether it is GPU-bound in FP32; we did not test that.
What GitHub-hosted runners cost (verified on docs.github.com)
GitHub's Actions runner pricing reference says GitHub rounds each job's minutes and partial minutes up to the nearest whole minute. The rates that matter for this comparison:
- Standard Linux 2-core x64: $0.006/min; Linux 1-core (
ubuntu-slim): $0.002/min; Linux 2-core arm64: $0.005/min. - x64 larger runners: 4-core $0.012, 8-core $0.022, 16-core $0.042, 32-core $0.082, 64-core $0.162, 96-core $0.252.
- arm64 larger runners: 4-core $0.008, 8-core $0.014, 16-core $0.026.
- GPU larger runner: Linux 4-core GPU $0.052/min.
Two plan rules change the math. First, standard GitHub-hosted runners are free and unlimited in public repositories, and there ubuntu-latest is a 4-core, 16 GB machine. Second, in private repositories ubuntu-latest is a 2-core, 8 GB machine that draws from your plan's monthly included compute, which GitHub now expresses as a dollar amount: $12 on GitHub Free, $18 on Pro and Team, $300 on Enterprise Cloud. Larger runners are only available on Team and Enterprise Cloud and are billed even for public repositories.
Side-by-side: StarSling vs GitHub runner options
| GitHub standard (private repo) | GitHub larger runners | StarSling usage-based | |
|---|---|---|---|
| 2 vCPU Linux x64 | $0.006/min (2-core, 8 GB) | $0.006/min (Advanced 2-core) | $0.004/min (8 GB) |
| 4 vCPU Linux | Not offered for private repos | $0.012/min x64, $0.008/min arm64 | $0.008/min x64 (16 GB) |
| GPU | None | T4 Linux 4-core $0.052/min | RTX 4090 $0.02522, RTX 5090 $0.03022, RTX PRO 6000 $0.06422/min |
| Billing unit | Whole minute per job | Whole minute per job | Whole second (new paid plans) |
| Free usage | $12-$18/month included compute; free in public repos | Draws from Team/Enterprise included compute; billed in public repos | One-time $8 (2,000 2-vCPU minutes) first month |
| OS | Linux, Windows, macOS | Linux, Windows, macOS | Ubuntu 26.04 and 24.04 only |
A month of agent-volume CI, priced
To make this concrete we priced a hypothetical month for a small team whose agents push a lot. The workload is made up for illustration; the rates are the live list prices above, and we assumed the same job duration on every runner, so no speed-up is credited to anyone.
- Workload A: 20,000 short agent PR checks of 40 seconds each. GitHub bills each as a full minute (20,000 minutes). StarSling's per-second billing counts 13,333 minutes.
- Workload B: 5,000 test jobs of 6 minutes 30 seconds. GitHub rounds each to 7 minutes (35,000 minutes); StarSling bills 32,500 minutes.
Monthly list-price totals for A plus B:
- GitHub 2-core x64 (private
ubuntu-latest): $330.00 - GitHub 4-core x64 larger runner: $660.00
- GitHub 4-core arm64 larger runner: $440.00
- StarSling 2 vCPU: $183.33
- StarSling 4 vCPU default: $366.67
For GPU work, 1,000 jobs of 10 minutes (10,000 minutes) cost $520.00 on GitHub's T4, $252.20 on StarSling's RTX 4090, $302.20 on the RTX 5090 and $642.20 on the RTX PRO 6000, again with equal minutes. These figures exclude GitHub's included compute, StarSling's one-time $8 allowance, storage and taxes. Our own table image with the full breakdown is among the official screenshots at the end of this post.
What the numbers tell you
- Short jobs are where per-second billing pays off. In workload A the 40-second checks lose a third of their billed time to GitHub's rounding. If your agents trigger many sub-minute jobs, this effect alone can be bigger than the rate difference.
- The default label is a size upgrade for private repos. StarSling recommends
starsling-ubuntu-26.04(4 vCPU) as the replacement forubuntu-latest. If you currently run the 2-coreubuntu-latestin a private repo, that swap raises the per-minute rate from $0.006 to $0.008. You only come out ahead if jobs get at least 25% shorter, or if per-second billing trims enough rounding. For a like-for-like move, usestarsling-ubuntu-26.04-2. - StarSling's own calculator assumes jobs take 1.7x longer on GitHub. Its estimate shows $204 vs $80 for 10,000 minutes on 4 vCPU because it bills GitHub for 17,000 minutes. That speed factor is the vendor's claim; our table leaves it out so you can add your own measured ratio.
- GitHub's arm64 larger runner matches the rate. If your build works on arm64, GitHub's 4-core arm64 runner is also $0.008/min, though it still rounds per job to the minute.
Editor's note
Good fit: teams with private repos, Linux-only workflows and a lot of short agent-triggered jobs, or teams already paying for GitHub 4-core and larger x64 runners. GPU test suites that today run on GitHub's T4 are also worth a trial, once StarSling opens GPU access for your org.
Not a fit: public open-source projects happy on free standard runners (StarSling minutes are never free after the first $8), and workflows that need Windows, macOS or Ubuntu 22.04.
Gotchas we found in the docs:
- The 2,000 free minutes are 2-vCPU-equivalent: on the default 4 vCPU runner they cover about 1,000 minutes.
- Older paid subscriptions keep whole-minute billing until migrated, so the per-second saving is not automatic for them.
- AI optimization PRs are included on paid plans but are not enabled by default for new accounts.
- GPU savings percentages assume FP32-bound speed-ups; per minute, the default RTX PRO 6000 costs more than GitHub's T4.
Common label errors and fixes (from StarSling's label reference): starsling-ubuntu-22.04 fails because only 26.04 and 24.04 exist; starsling-ubuntu-latest fails, use starsling-ubuntu-26.04; starsling/ubuntu-24.04 fails because the base label uses hyphens and / only introduces a selector; unlisted GPUs such as gpus=rtx-3090:1 do not match any runner.
How to decide in an afternoon
- Export one month of Actions usage from your GitHub billing page and note which runner sizes and how many jobs you ran.
- Count how many jobs finish in under a minute; that is the share where per-second billing helps most.
- Move one busy workflow to the StarSling label of the same size and compare job durations over a week.
- Only then decide whether the bigger default runner, or a GPU runner, is worth its rate.
If you also run agent code in sandboxes, our Agent37 vs Vercel Sandbox vs E2B cost comparison covers the other half of the compute bill. For GPU work outside CI, see Modal vs Runpod vs Lambda: what an H100 hour really costs, and for giving agents command-line tools in the first place, the CLI-Anything tutorial.
FAQ
Is StarSling free?
Only at the start. New accounts get 2,000 2-vCPU-equivalent minutes ($8 of compute) in the first month, with no credit card required. After that every minute is paid; the open-source skills such as ci-speedup and ci-score are free and run with your own agent.
Is StarSling really 33% cheaper than GitHub Actions?
Per minute, yes against GitHub's x64 larger runners at 2 and 4 vCPU; the gap falls to 21% at 64 vCPU. Against a private-repo ubuntu-latest (2-core, $0.006) the StarSling default 4 vCPU runner ($0.008) costs more per minute, and against free public-repo runners it always costs more.
Do I have to rewrite my workflows?
No. StarSling runs existing GitHub Actions workflows; you change the runs-on label after installing its GitHub App on your organization.
Pick the runner size that matches what you run today, measure a week, and let real job times rather than headline percentages decide.
Official screenshots at the end of this post: StarSling pricing (starsling.dev), GitHub Docs (docs.github.com); last table by lightmind24.
Last verified: 2026-11-19
Sources: StarSling pricing, StarSling docs: Pricing, StarSling docs: Label reference, GitHub Docs: Actions runner pricing, GitHub Docs: GitHub Actions billing, GitHub Docs: GitHub-hosted runners.
lightmind24 is an independent review site, not affiliated with StarSling or GitHub.
댓글
댓글 쓰기