A Runtime sandbox bills the way serverless does. It charges for CPU only while it is busy. It is a whole Linux machine, though, and it keeps its memory, files and processes between calls. While a sandbox waits on a model, a 2 vCPU, 4 GiB sandbox costs $0.03125 an hour. After 300 seconds with nothing happening, it pauses itself. The next request wakes it (pause when idle).
Where the term comes from
The CNCF Serverless Working Group's whitepaper defines serverless computing as "the concept of building and running applications that do not require server management". Code is packaged as functions. The functions are "executed, scaled, and billed in response to the exact demand needed at the moment". There is "no charge when code is not running". AWS Lambda is the reference example. It bills duration by the millisecond, and it charges a price per million requests.
What serverless gives up
| Need | A function | A sandbox on Runtime |
|---|---|---|
| Run untrusted, generated code | Usually refused, or one language | Any language, sudo, its own kernel |
| Keep state between calls | Reload it each time | Files, memory and processes stay |
| Long jobs | Capped per call | No time limit: runs while it works, while credit lasts |
| A shell, a desktop, a port to open | No | Yes |
| Pay nothing while idle | Yes | Paused storage only, $0.08 per GB a month |
This gap is why agent platforms run tools in sandboxes instead of functions. An agent's work is stateful. It runs code that nobody has reviewed. It also waits long stretches between steps.
How Runtime gets close to scale-to-zero
- CPU is measured. An idle sandbox pays for its memory and a floor of a twentieth of a vCPU. It does not pay for every vCPU it holds (CPU floor).
- Idle pause. After 300 seconds with no request, command, connection, traffic or CPU use, the sandbox pauses. It then stops paying for compute.
- Wake on request. The next command, file call or preview visit wakes the sandbox.
Related
Sources
Checked 25 October 2026.