Sandbox leaderboard

How fast you get a sandbox, a command, a coding agent's first words and 100 sandboxes at once, beside 28 other clouds.

As of 2026-11-01: a sandbox in 1.3 s, its first command 174 ms later. Factory Droid writes its first words 4.1 s after create. 100 at once: Tangle is #28 of 29 on ComputeSDK's board, at a 27 s median.

First agent words

Seconds from create to a coding agent's first text, 1 run each. Lower is better.

Factory Droid: 4.1 s4.1Factory DroidCodex: 4.1 s4.1CodexPrime Agent: 4.2 s4.2Prime AgentClaude Code: 5.4 s5.4Claude CodeOpenCode: 5.7 s5.7OpenCodePi: 6.3 s6.3PiKimi Code: 7.9 s7.9Kimi CodeACP: 11 s11ACPHermes: 17 s17HermesOpenClaw: 23 s23OpenClaw

Every verb

Median seconds per call on one sandbox, 3 runs. Lower is better.

create: 946 ms0.95createsnapshot: 1.4 s1.4snapshotstop: 1.0 s1.0stopresume: 854 ms0.85resumerestore: 829 ms0.83restorebranch a child: 825 ms0.82branch a childdelete the child: 2.7 s2.7delete the childdelete the snapshot: 868 ms0.87delete the snapshotdelete: 3.4 s3.4delete

100 at once

Median and p95 time to interactive on a log scale. Lower left is better.

10 ms10 ms100 ms100 ms1.0 s1.0 s10 s10 s100 s100 smedian →p95 ↑Isorun: 49 ms median, 57 ms p95IsorunPromigence: 65 ms median, 82 ms p95Miosa: 67 ms median, 74 ms p95Gravixlayer: 102 ms median, 107 ms p95Createos: 136 ms median, 176 ms p95Givemeanode: 322 ms median, 467 ms p95Arker: 324 ms median, 394 ms p95Superserve: 371 ms median, 655 ms p95Blaxel: 380 ms median, 425 ms p95Vercel: 423 ms median, 562 ms p95Beam: 548 ms median, 732 ms p95Daytona: 668 ms median, 747 ms p95Modal: 751 ms median, 972 ms p95Cloudflare: 779 ms median, 1.2 s p95Mosaic: 889 ms median, 3.5 s p95Buddy: 905 ms median, 1.1 s p95Tensorlake: 1.0 s median, 2.2 s p95E2B: 1.3 s median, 1.5 s p95Novita: 1.4 s median, 2.0 s p95Cloud Run: 1.7 s median, 3.1 s p95Upstash: 1.9 s median, 3.8 s p95Tenki: 2.1 s median, 2.2 s p95Sail: 2.1 s median, 2.2 s p95Freestyle: 2.8 s median, 5.2 s p95Microsandbox: 4.7 s median, 5.2 s p95Tangle, 2026-10-22, serving 45fc1ecd: 6.8 s median, 15 s p95, failed a checkSandbox0: 10 s median, 23 s p95Codesandbox: 16 s median, 48 s p95Tangle, 2026-10-31, serving 99d6aca1: 27 s median, 52 s p95TangleRun Cloud: 36 s median, 92 s p95

Coding agents: create to first text

run full-20261101212030451-6c7376b2, 2026-11-01, serving d8ed76cabf1f. One prompt on one fresh sandbox per agent.

#agentp50n
1Factory Droid4.1 s1
2Codex4.1 s1
3Prime Agent4.2 s1
4Claude Code5.4 s1
5OpenCode5.7 s1
6Pi6.3 s1
7Kimi Code7.9 s1
8ACP11 s1
9Hermes17 s1
10OpenClaw23 s1

Clouds: 100 sandboxes at once

ComputeSDK's burst board read 2026-10-30; Tangle rows by the same method on production.

#cloudmedianp95reachedn
1Isorun49 ms57 ms100%100
2Promigence65 ms82 ms100%100
3Miosa67 ms74 ms100%100
4Gravixlayer102 ms107 ms100%100
5Createos136 ms176 ms100%100
6Givemeanode322 ms467 ms100%100
7Arker324 ms394 ms100%100
8Superserve371 ms655 ms100%100
9Blaxel380 ms425 ms100%100
10Vercel423 ms562 ms100%100
11Beam548 ms732 ms100%100
12Daytona668 ms747 ms100%100
13Modal751 ms972 ms100%100
14Cloudflare779 ms1.2 s100%100
15Mosaic889 ms3.5 s100%100
16Buddy905 ms1.1 s100%100
17Tensorlake1.0 s2.2 s100%100
18E2B1.3 s1.5 s100%100
19Novita1.4 s2.0 s100%100
20Cloud Run1.7 s3.1 s100%100
21Upstash1.9 s3.8 s100%100
22Tenki2.1 s2.2 s100%100
23Sail2.1 s2.2 s92%100
24Freestyle2.8 s5.2 s100%100
25Microsandbox4.7 s5.2 s100%100
26Tangle2026-10-22, serving 45fc1ecd✱6.8 s15 snot all100
27Sandbox010 s23 s100%100
28Codesandbox16 s48 s100%100
29Tangle2026-10-31, serving 99d6aca127 s52 s100%100
30Run Cloud36 s92 s95%100

Each row creates 100 sandboxes at once and reports the median and p95 time to interactive. Failed checks on the Tangle rows: tracked_resources_settled, first_token_observed, error without a code, branch_n16_children_deleted, burst_matched_n100_all_reached_first_command, burst_matched_n100_cleanup, burst_n25_cleanup, matrix_burst-100_all_reached_first_command, matrix_burst-100_cleanup, matrix_burst-25_cleanup, matrix_burst-25-unmatched_all_reached_first_command, matrix_restore_all_reached_first_command, matrix_restore_integrity.

One sandbox, every verb

run full-20261101212030451-6c7376b2, 2026-11-01, serving c4e7406e47cc. Created, snapshotted, stopped, resumed, restored, branched, then everything deleted.

verbp50n
create946 ms3
snapshot1.4 s3
stop1.0 s3
resume854 ms3
restore829 ms3
branch a child825 ms3
delete the child2.7 s3
delete the snapshot868 ms3
delete3.4 s3
The same journey on the desktop shape
verbp50n
create802 ms3
snapshot1.4 s3
stop1.3 s3
resume2.0 s3
restore1.5 s3
branch a child1.5 s3
delete the child2.8 s3
delete the snapshot560 ms3
delete4.1 s3

Many at once

run full-20261031092953471-99d3d630, 2026-10-31, serving 99d6aca10260; branches from run full-20261031092953471-99d3d630, 2026-10-31, serving 99d6aca10260.

creates in a burstp50p90n
10 creates at once4.3 s4.9 s10
25 creates at once8.0 s11 s25

Each sandbox timed to its first command; n = sandboxes that reached one.

branch or forkp50p95n
1 child1.1 s1.5 s20
4 children1.7 s2.9 s20
16 children3.8 s7.1 s20
fork from a snapshot1.8 s3.3 s24
fork from a checkpoint2.2 s3.1 s20

Branch: the call to the slowest child's first command. Fork: a new sandbox from a saved snapshot or a running checkpoint.

First Router call from a fresh sandbox

run full-20261101212030451-6c7376b2, 2026-11-01, serving c4e7406e47cc. Each sandbox carries its own key; the path is how the command reached it.

#pathp50n
1exec1.2 s1
2terminal shell2.4 s1
3direct to the runtime938 ms1
4startup script5.4 s1