Compare · CI runnersSandal vs Depot
Depot’s Docker build caching is the deepest in the market, on a bare-metal plus NVMe design built for that one problem. Sandal runs the same class of hardware for every job, prices it lower per minute, bills by the second, and puts sandboxes and mobile simulators on the same account.
- $ / min · Linux 2 vCPU
- 0.004Depot0.003Sandal
- $ / min · macOS
- 0.080Depot0.060Sandal
- Billing granularity
- MinuteDepotSecondSandal
- Also on the account
- SandboxesDepotSandboxes · simulatorsSandal
The honest splitTwo good answers. Here is which one is yours.
- Most of your CI minutes go to building and pushing container images.
- Cache hit rate on Docker layers is the number that decides your bill.
- You want Depot’s build API and Connect integration as the centre of the pipeline.
- Your CI is tests and builds across Linux, Windows and macOS, not primarily image builds.
- You want per-second billing and a lower published rate on the same hardware class.
- The same team needs agent sandboxes and iOS simulators from one vendor.
Many teams run both. runs-on is per job and a sandbox is per task; nothing here is all or nothing.
Side by sideWhat each one actually ships.
Documented capability as of August 2026. Where the honest answer is a qualifier rather than yes or no, the qualifier is printed. Scroll sideways on a small screen.
| Area | Depot | Sandal | Why it matters |
|---|---|---|---|
| Runtime | |||
| Hardware | Bare metal with NVMe, designed around build caching. | Dedicated cores on 4.5 GHz EPYC with local NVMe. | Both leave shared hyperthreads behind. |
| Cache | Docker layer cache that is the deepest in the market. | Sticky disks: named volumes restored just in time, no cap. | Image-build pipelines versus everything else. |
| Isolation | |||
| Tenancy | Ephemeral runner per job. | Single-tenant microVM per job, destroyed on exit. | A fork PR cannot touch your release job. |
| Pricing | |||
| Billing | Per minute. | Per second; queue time free. | Short jobs pay for seconds, not minutes. |
| Surfaces | |||
| Operating systems | Linux, Windows, macOS. | Linux, Windows, macOS and native ARM. | Parity on OS coverage. |
| Beyond CI | Sandboxes and a build API. | Sandboxes, iOS and Android simulators, Xcode builds, Windows and macOS machines. | What else the account can run. |
Capabilities change and rates change. If something here is out of date, tell us and we will fix it the same week.
PriceSame shape. Two meters.
Published rates, August 2026. macOS on Apple silicon: Depot $0.080, Sandal $0.060 per minute. Depot bills by the minute; Sandal by the second.
- DepotLinux · 2 vCPU
- 0.004$ / minute
- SandalLinux · 2 vCPU · paused = $0
- 0.003$ / minute
Five real job lengths on the published 2 vCPU Linux rates. A 71-second job bills as two minutes on a per-minute meter; times thousands of jobs a month, that is the difference.
MigrationOne label changes.
Sandal registers as a self-hosted runner, so GitHub keeps orchestrating. Every action, cache, secret and reusable workflow keeps working.
- 01≈ 2 min
Install the app
Grant Actions and workflow read on the repos you want to move.
- 02≈ 5 min
Swap the label
npx sandal migrate rewrites runs-on and opens a PR. Keep depot/build-push-action jobs on Depot if the layer cache is the point.
- 03≈ 1 week
Watch a week of jobs
Compare wall-clock time and the two invoices, then decide per job.
Where Depot is aheadWhat we would rather you read here than find later.
A comparison page where the vendor wins every row is a brochure. These are real, and we would rather you land on the right tool than land on us and leave in a month.
- 01
Container image builds.
If most of your CI minutes go to building and pushing images, Depot’s layer cache decides the bill and it is the deepest in the market. Keep those jobs there.
Questions teams ask before moving off Depot.
Yes. runs-on is per job, so image builds can stay on Depot while tests and everything else move. Both register as self-hosted runners; GitHub orchestrates.
More comparisonsThe rest of the field in ci runners.
Run the same job twice.
Point one workload at Sandal and one at Depot. The numbers on this page are ours; the ones that matter are yours.