Pricing

Alien provisions Sandboxes in your customer's cloud. The customer pays the cloud provider for the underlying compute, and Alien charges a small management fee. See Pricing for Alien's current rates.

Prices shown for US East regions. Last verified: August 2026.

Sandbox cost varies by more than an order of magnitude across platforms, and the differences are structural rather than a matter of rates. This page names the charges that do not appear on the line you expect.

AWS

Per-second compute while a MicroVM runs, on the image's baseline memory. A running MicroVM auto-scales to 4x baseline, and the memory and vCPU above baseline are billed for the duration they are used. Size on baseline and you can be charged up to four times it.

Then the parts that are easy to miss. These snapshots are internal — the build-time image snapshot and suspend/resume state, not a user-callable feature; snapshot() is unavailable on every platform.

• Snapshot storage, continuous.
• Snapshot read on every start and resume.
• Snapshot write on every suspend.
• A one-week minimum retention on image storage. An image you create and delete an hour later still bills for a week. That matters if you roll images often — each roll starts its own week, and an image that fails to delete keeps billing whether or not you noticed.

GCP

No additional charge. Sandboxes run as subprocesses inside the hosting Cloud Run service's existing CPU and memory. The cost is the headroom you provision on that service.

This is the cheapest option by a wide margin and the reason GCP is worth considering at all, given how much it cannot do — no reconnection, no preview, no enforced limits. Size the hosting service for your app plus its concurrent sandboxes; nothing else appears on the bill.

Azure

Container Apps Sandboxes bill per second, only while a sandbox runs, on the Consumption plan's vCPU and memory rates across resource tiers from XS to XL. Stopped sandboxes accrue no CPU or memory charge, and the sandbox group itself is free — it carries no SKU and no capacity. Snapshots are free during the preview and move to Blob Storage rates afterwards.

This is the most forgiving shape for short, frequent turns of any platform that charges at all. Two caveats: Sandboxes are in public preview, so the rates can move before GA; and the scale-to-zero claim is Microsoft's published word rather than something measured here.

Kubernetes

The cluster and its gVisor node pool, standing. Alien adds a warm pool of idle pods, because cold start measured 79s against 2.7s warm and 79s per agent turn is unusable — those idle pods occupy nodes you are already paying for.

Local

Your machine.

What a per-turn agent workload actually costs

Take a 20-second turn — the shape of an agent running generated code.

Read the table as turns, not sessions. One conversation reusing a session across many turns costs about one session's runtime, not one session per turn. The numbers above assume each turn's work runs back to back; concurrent conversations multiply them.

Both per-second platforms track execution rather than readiness, so the shape of your workload matters less than it looks. AWS is predictable, but the snapshot I/O and the one-week image retention are real and neither shows up where you look first. On Azure the sandbox group is free and stopped sandboxes cost nothing, which suits bursty turns — at the price of the thinnest capability set of any platform: no file transfer, no enforced ceilings, no enforced deny.

GCP is close to free and correspondingly limited. If your workload fits inside a single turn, that trade is very good. If it does not, GCP is not cheaper — it is unusable, and cost is the wrong axis to have chosen it on.

Sources

• AWS Lambda Pricing — MicroVM compute, snapshot I/O, image retention
• Azure Container Apps Pricing — Sandboxes follow Consumption-plan rates
• Azure Container Apps Sandboxes overview
• Google Cloud Run Pricing — the hosting service a GCP sandbox runs inside