Skip to content
Runtime

Agent sandboxes.
Up to 77% cheaper.

A full Linux environment for your agents.

Try for free

20 free hours. No credit card.

1,000 sandbox hours

Estimated compute cost

Comparison details

2 vCPUs · 4 GiB · 20 CPU-seconds per minute

Calculate your savings

Let your agent get to work.

Build applications.

Analyze data.

Research the web.

Run in the background.

RunningComplete

Stay isolated.

Illustrative examples.

Fast to start.

Median time to first command
454 ms
Median creation
270 ms
95th percentile to first command
480 ms
  • 50 sequential starts
  • Zero failures
  • Trial template benchmark
Benchmark details

Measured on 22 September 2026 at 01:46 UTC, on one AMD EPYC 7371 host running Firecracker 1.17.0 on kernel 6.1.186. Each sandbox had the trial shape: 2 vCPUs, 4096 MiB of memory with a balloon, and a 4 GiB disk. Every start came from the retained trial template, 50 of them, one after another.

Creation ends when the sandbox reports running. Time to first command measures from the start of creation until an echo command completes, which is what makes a sandbox usable here. The median is the middle result; the 95th percentile describes the slower end of these 50 runs. Displayed times are rounded to the nearest millisecond.

These are host-local measurements taken through the host driver for the trial shape, not current default startup times or public-API latency. They exclude the trip over the public network: one run through the public API the same night reached running in about 0.8 s and its first command in about 1.1 s, which is a single run rather than a percentile. These figures do not establish performance under concurrent load.

One sandbox.
Full control.

Explore the sandbox
ExecuteRun commands
FilesRead and write
NetworkReach the web
StateSave and resume

Illustrative workflow: execute a command, save its output to a file, fetch a public web page, and save files and memory for the next run.

Right where you left it.

Pause the work. Keep the state.

Files and memory stay with you.

01RunningWork in motion
02PausedState preserved
03ResumedSame place. Keep going.
Illustrative demo. Explore pause and resumePaused storage is billed separately. Retention limits apply.

Its own computer.
Your boundaries.

Explore security

Isolated machines

Scoped access

AllowedBlocked

Controlled network

Your compute bill. 77% smaller.

Runtime $639 per month. E2B and Daytona each $2,760 per month. Estimated compute savings 77%.
Between 10,000 and 10,000,000 runs per month.
Workload details
Runtime active CPU
$139
Reserved memory
$500

One workload, compared at published compute rates: 60 seconds per run and 20 total CPU-seconds across two vCPUs, with 4 GiB of memory. Runtime charges for active CPU, with a 0.25-vCPU billing floor. E2B and Daytona charge for allocated CPU. Equal CPU counts do not establish equal performance.

E2B and Daytona have the same compute rate for this configuration; the combined bar shows the cost of either provider, not their sum. Estimates exclude plan fees, creation fees, storage, networking, taxes, retries, free credits and negotiated discounts. Paused storage is billed separately. Values are rounded; savings use unrounded costs.

Compute rates checked September 21, 2026: Runtime, E2B and Daytona.

1 min / run · 17% CPU · 2 vCPUs · 4 GiB

Estimated compute only. Other charges apply.

Estimated monthly compute: Runtime $639. E2B and Daytona each $2,760. 77% lower compute cost.

Know what you’re building on.

What does each sandbox include?

An isolated Linux environment for commands, packages, files, and background processes. Each sandbox runs in its own Firecracker microVM.

Can I keep my current agent and model?

Yes. Keep your model, prompts, and application logic. Connect through Python, JavaScript, CLI, API, or MCP, and check the features your application needs. Explore integrations.

What is saved when I pause?

Files and memory are saved so the sandbox can continue from its last state. Paid paused storage is billed separately; retention limits and capacity on wake apply. Read the pause and storage terms.

How does compute billing work?

Runtime meters active CPU and reserved memory separately, with a CPU billing floor. The calculator shows estimated compute costs; storage, networking, plan fees, taxes, credits, and other charges are excluded. Compare your workload.

What network access is available?

Sandboxes reach the public web on ports 80 and 443; private and internal addresses are refused. Accounts that have bought credit can open other host:port pairs, such as a database. Nothing connects in except through a preview: an HTTPS address for one port, private by default. Read the network details.

What is included in the free trial?

20 free hours of sandbox compute with no credit card. Run up to three trial sandboxes at once, each up to 2 vCPUs and 4 GiB of memory; idle running time counts. Model calls are separate. After the trial, add prepaid credit by card from $10; a trial sandbox never spends it. See trial details.

Explore the docs

Let your agent
handle the setup.

Sign up, then paste the setup prompt into Claude Code, Codex, or your coding agent. It handles setup; you approve the connection.

20 free hours. No credit card.

Try for free