Cloud sandboxes, measured

Oct 22–26, 2026 · 8 providers (9 configs) · 2 caller locations: Florida, US & Frankfurt, DE · 100 runs per main test, 20 per location test

Oct 22–26, 2026 · 8 providers (9 configs) · 2 caller locations: Florida, US & Frankfurt, DE · 100 runs per main test, 20 per location test

By location

How network distance changes latency: the same tests run from EU-Central (Frankfurt) and US-East (Florida).

Cold start by caller location

Median time from each caller location · Colored bar marks the faster location · Lower is better

Oct 26, 2026 · 5–20 runs per test

E2BSandbox in Oregon
EU-Central
0.67 s
US-East
0.48 s
×1.4 faster from US-East
DaytonaAPI in Virginia
EU-Central
1.5 s
US-East
0.99 s
×1.5 faster from US-East
Upstash BoxSandbox in Virginia
EU-Central
3.0 s
US-East
2.5 s
×1.2 faster from US-East
BoatSandbox in Northern France
EU-Central
0.78 s
US-East
1.1 s
×1.5 faster from EU-Central
boxdSandbox in Frankfurt
EU-Central
1.6 s
US-East
1.0 s
×1.6 faster from US-East
FreestyleSandbox in SF Bay Area
EU-Central
1.1 s
US-East
0.81 s
×1.3 faster from US-East
TenkiSandbox in Seattle
EU-Central
1.1 s
US-East
0.91 s
×1.2 faster from US-East
SuperserveSandbox in Virginia
EU-Central
0.71 s
US-East
0.49 s
×1.5 faster from US-East

Run a command

fastest location
E2B
0.12 sUS-East×1.5
Daytona
0.10 sUS-East×1.6
Upstash Box
0.22 sUS-East×2.5
Boat
0.11 sEU-Central×2.9
boxd
0.17 sEU-Central×1.7
Freestyle
0.16 sUS-East×1.6
Tenki
0.17 sUS-East×1.4
Superserve
0.15 sUS-East×1.6

Write + read 10 MB

fastest location
E2B
1.3 sUS-East×2.1
Daytona
1.8 sUS-East×1.1
Upstash Box
2.1 sUS-East×1.9
Boat
3.5 sEU-Central×2.0
boxd
0.74 sEU-Central×3.9
Freestyle
2.1 sUS-East×2.3
Tenki
2.0 sUS-East×1.5
Superserve
0.61 sUS-East×3.6

Agent tool replay

fastest location
E2B
14 sUS-East×1.6
Daytona
17 sUS-East×1.3
Upstash Box
21 sUS-East×1.8
Boat
23 sEU-Central×1.8
boxd
17 sEU-Central×1.8
Freestyle
19 sUS-East×1.7
Tenki
15 sUS-East×1.6
Superserve
11 sUS-East×2.0

Boat (Northern France) and boxd (Frankfurt) run their sandboxes in Europe; the remaining providers run in US-East. In-sandbox CPU, memory, disk, and package install speed do not change with caller location and are omitted here.

Where the time goes

Step by step, from median timings, in each provider's own time.

TraceCold start · every provider, median of each step
Fastest
519ms
Superserve · provider time
Slowest
2.62s
Upstash Box · provider time
Spread
5.1×
slowest vs fastest
Failures
0
of 900 attempts
Provider
0s1s2s3s
create()First commandRanked by provider time · pick a provider for its full trace.
TraceCold start · every provider, median of each step
Fastest
495ms
E2B · provider time
Slowest
2.62s
Upstash Box · provider time
Spread
5.3×
slowest vs fastest
Failures
0
of 900 attempts
Provider
0s1s2s3s
create()First commandOurs: Sandbox SDKRanked by provider time · pick a provider for its full trace.

Start-up

Cold start is the wait before an agent can do anything: from create() until the first command runs, the provider's own time.

Cold start

create() to first command, without our SDK's setup · Median seconds · Lower is better

Cold start

create() to first command, without our SDK's setup · Median seconds · Lower is better

Agent work

Real coding-agent tasks, so CPU, disk and network count as much as the API.

Agent tool replay

100 tool calls in a row: write, read, edit, list, grep, run a test · Median seconds · Lower is better

Agent tool replay

100 tool calls in a row: write, read, edit, list, grep, run a test · Median seconds · Lower is better

Everyday calls

What an agent does hundreds of times per task: run a command, move files, reattach to a sandbox.

Run a command

A no-op command, mostly network and API overhead · Median seconds · Lower is better