# Agent sandboxes compared: start time and price

Code and browser sandboxes start in about a second or less, cloud Android phones in 1 to 23 seconds. Published and measured start times with per-minute prices.

Comparisons4 min read

In this article

Code sandboxes start in about half a second to a second, browser sessions in a quarter of a second to just over a second, and cloud Android phones ([mobile sandboxes](https://mobilebox.dev/mobile-sandbox)) in one to twenty-three seconds. Mobilebox runs cloud Android phones for AI agents at $0.01 a minute, ready for a first screenshot in 19.4 s in our November 8, 2026 measurement, while a minute of a default code or browser sandbox costs a tenth of a cent or less and a minute of a cloud phone one to six cents.

The code and browser numbers below are published by vendors or by third-party benchmarks and are not measured by us; the Android numbers are [our own measurements](https://mobilebox.dev/benchmarks) from November 8, 2026.

All prices are list prices for one running sandbox of the default or smallest size, read on November 8, 2026. Start time means "create" until the sandbox can run its first command, load its first page or return its first screenshot, depending on the kind.

## Code sandboxes

| Sandbox | Start time | Source | Price per minute (default size) |
| --- | --- | --- | --- |
| Vercel Sandbox | 0.53 s median | ComputeSDK, Nov 6, 2026 | $0.0014 idle, up to $0.0057 at full CPU (pricing) |
| Daytona | 0.60 s median; vendor: "Sub 90ms sandbox creation" | ComputeSDK; Daytona | $0.0011 (pricing) |
| Cloudflare Sandbox | 0.67 s median; vendor: "often … in the 1-3 second range" | ComputeSDK; Cloudflare docs | from $0.00005, needs the $5/month Workers Paid plan |
| Modal Sandboxes | 0.80 s median; vendor: "start in under one second" | ComputeSDK; Modal | from $0.00035 at the smallest request |
| E2B | 1.13 s median; vendor: resume "approximately 1 second" | ComputeSDK; E2B docs | $0.0018 to 0.0028 (pricing) |

ComputeSDK creates 100 sandboxes at once per provider and measures until the first command succeeds, network included. Its rankings move from run to run: in an August run, [as reported by Bex](https://bex.co/blog/2026/10/04/superagent-sandbox-benchmark-cold-start-concurrency), Daytona was fastest but succeeded on only 37% of starts. Read a single median as "under a second", not as a ranking.

## Browser sandboxes

| Browser | Start time | Source | Price per minute |
| --- | --- | --- | --- |
| Kernel | 249 ms median | Browser Arena, Nov 8, 2026 | $0.0010 headless (pricing) |
| Browser Use Cloud | 391 ms median | Browser Arena | $0.00033 (pricing) |
| Browserbase | 790 ms median | Browser Arena | $0.0020 on Developer (pricing) |
| Steel | 1,186 ms median; vendor: "Start a session in <1s" | Browser Arena; Steel | $0.0017 (pricing) |
| Hyperbrowser | vendor: "<500ms" from request to a ready browser | Hyperbrowser | $0.0017 |

Every browser benchmark we found is run by a vendor or likely affiliated with one. Steel’s [browserbench](https://github.com/steel-dev/browserbench), for example, measures a full cycle (create, connect, load a page, release) and has reported different orders between runs.

## Cloud Android phones

| Phone | Cold start to first screenshot (median) | Price per minute | Source |
| --- | --- | --- | --- |
| Limrun | 0.98 s | $0.04 on Entry, $0.06 trial overage | measured |
| Mobilebox | 19.4 s | $0.01 | measured |
| Phone Harness | 23.0 s (runs clustered at ~21 s and ~67 s) | $0.05 | measured |
| Mobilerun | not measured | $0.03 overage | pricing |

A phone is a whole operating system with an app install, so it is slower to start and costs more per minute than a process or a browser tab. The spread between phones is also wide: some services keep phones booted and hand one over in about a second, others boot one per request. Our [benchmark page](https://mobilebox.dev/benchmarks) has the method, p90s, tap latency, install times and the raw CSV.

## What start time means for your agent

- **Short tasks** pay start time on every run. A one-minute check on a phone that takes 20 seconds to start spends a quarter of its wall clock waiting.

- **Long sessions** barely notice it. Price per minute and what a stopped sandbox keeps matter more.

- **Parallel runs** need the vendor’s concurrency on your plan, not only a fast single start.

Pick the smallest surface that does the job: [a process, a browser or a phone](https://mobilebox.dev/blog/native-app-or-browser). Then compare start time and price within that kind, not across kinds.

## Try it in your agent in 60 seconds

Paste this into Claude Code, Cursor, Codex or any agent that can read a web page. It asks for your email, gets its own key at once and adds the phone tools; the 6-digit code from your inbox keeps the workspace.

```
Set up Mobilebox for me: read https://mobilebox.dev/setup.md?via=agent-sandbox-start-times and follow it.
```

[What your agent reads](https://mobilebox.dev/setup.md?via=agent-sandbox-start-times)

In this article

[All articles](https://mobilebox.dev/blog)

## Next questions

- [Cloud phone or a local emulator?](https://mobilebox.dev/blog/mobilebox-vs-local-emulator?via=next-emulator)

- [Mobilebox, Limrun or Mobilerun?](https://mobilebox.dev/blog/mobilebox-vs-limrun?via=next-limrun)

- [How do I add it to Cursor, Codex or VS Code?](https://mobilebox.dev/blog/android-mcp-cursor-codex-vscode?via=next-setup)

- [Can my agent run 50 phones in parallel?](https://mobilebox.dev/blog/run-android-phones-in-parallel?via=next-parallel)

- [Can my agent use an iPhone?](https://mobilebox.dev/ios?via=next-iphone)

- [Can each phone get a residential IP?](https://mobilebox.dev/blog/residential-ip-per-phone?via=next-residential-ip)

- [Does my agent stay logged in between runs?](https://mobilebox.dev/blog/park-resume-android?via=next-logins)

- [Can I watch my agent work live?](https://mobilebox.dev/live-view?via=next-live)