Runtime

GitHub Codespaces vs an agent sandbox: which one should run AI agents?

Codespaces gives a person a cloud dev environment per repository; an agent sandbox gives code a machine it creates and drives by API.

Runtime charges $0.03125 to $0.08 an hour for 2 vCPUs and 4 GiB, against $0.18 an hour for a 2-core codespace, because it bills the CPU the code uses rather than the machine's cores. A new sandbox ran its first Python command 351 ms after the create request at the median, measured on 24 September 2026 (speed), and every command is one API call. GitHub rates checked 25 September 2026.

What Codespaces is built for

GitHub describes a codespace as a development environment for a repository. When you create one, "a virtual machine (VM) is created", "dedicated and private to you", and "GitHub Codespaces uses a Docker container as the development environment". You connect to it with an editor, or from a terminal with gh codespace ssh.

  • Files. /workspaces is "a persistent directory that is mounted into the container". Changes elsewhere survive a stop and start but "are not preserved when you rebuild the container".
  • Idle timeout. "By default this period is 30 minutes", and it can be set from 5 to 240 minutes.
  • Retention. Stopped codespaces are deleted after 30 days inactive by default, and 30 days is the maximum.
  • API. The REST API creates a codespace in a repository and starts and stops it (POST /repos/{owner}/{repo}/codespaces, POST /user/codespaces/{codespace_name}/start). Commands inside run over a connection such as gh codespace ssh.

Side by side

GitHub Codespaces Runtime sandbox
Made for A developer working on one repository Code that an agent or a program starts, runs and throws away
Machine A dedicated VM with the dev container inside A Firecracker microVM with its own Linux kernel
Run a command Connect, then run it sbx.exec(...), with exit code, output and a timeout in the result
Sizes 2 to 32 cores vCPUs, memory and disk set per sandbox; 2 vCPU and 4 GiB by default
Idle Stops after 5 to 240 minutes; files kept Pauses with files, memory and running processes, kept 1 to 365 days
Copies New codespaces from the repository Fork a running sandbox, memory included, into 1 to 10 copies
Editor An editor, or gh codespace ssh VS Code and JetBrains over SSH (runtime sandbox ssh config --install)
Billing By machine type while running, plus storage Measured CPU plus reserved memory while running; paused storage

What each costs

GitHub's rates, checked 25 September 2026: compute from $0.18 an hour on a 2-core machine to $2.88 on 32 cores, "proportional to the number of processor cores", and storage at $0.07 per GB-month. A personal account on GitHub Free includes 120 hours of compute and 15 GB-month of storage a month; GitHub Pro includes 180 hours and 20 GB-month.

Runtime at 2 vCPU and 4 GiB: $0.03125 an hour waiting (memory plus a CPU floor of a twentieth of a vCPU) and $0.08 an hour with both CPUs busy.

Usage, past any free allowance Codespaces, 2-core Runtime, 2 vCPU and 4 GiB
One hour $0.18 $0.03125 waiting, $0.08 busy
A working month, 8 hours × 21 days 168 × $0.18 = $30.24 168 × $0.03125 = $5.25 to 168 × $0.08 = $13.44
Stored while idle $0.07 per GB-month $0.08 per decimal GB-month, memory included

An agent spends most of an hour waiting on its model, which is where measured CPU pays: the sandbox is charged for the floor, not for two idle cores (pricing).

The Runtime equivalent

A named environment per task or per user: found again if it exists, paused after 15 quiet minutes, and woken by the next command.

TypeScriptimport { Sandbox } from "withruntime";const dev = await Sandbox.getOrCreate("issue-4821", { idlePauseSeconds: 900 });await dev.exec("git clone https://github.com/octocat/Hello-World.git repo || true", {  timeoutMs: 120_000,});const run = await dev.exec("cd repo && git log --oneline -1");console.log(run.stdout);
Pythonfrom withruntime import Sandboxdev = Sandbox.get_or_create("issue-4821", idle_pause_seconds=900)dev.exec("git clone https://github.com/octocat/Hello-World.git repo || true", timeout_ms=120_000)print(dev.exec("cd repo && git log --oneline -1").stdout)

To look over the agent's shoulder in your own editor:

Terminalruntime sandbox ssh config --install   # then connect VS Code to issue-4821.runtime

Which one fits

  • Pick Codespaces for people: a developer, a class or a reviewer who wants a ready editor for a repository, inside GitHub's billing and its free monthly hours.
  • Pick an agent sandbox for programs: an agent that needs a fresh machine per task, a hundred at once, commands with structured results, network rules it cannot change, keys it never holds, and a bill that stays small while it waits on the model.

A paid Runtime account runs 100 sandboxes at once to start, and new accounts get 50 free sandbox hours, no card.

More: coding agent sandbox, per-user dev environments, pause and resume a sandbox, SSH and editors.

Sources

Checked 25 September 2026.

Facts on this page were checked on 25 September 2026.

Claim 50 hours free