P95 stands for the 95th percentile. 95% of iterations completed at or below this time. Tells you the 'typical worst case' — what a user experiences on a bad-but-not-extreme run.

P99 stands for the 99th percentile. 99% of iterations completed at or below this time. It highlights the tail latency (the rare outlier spikes). With 100 iterations, this is the 99th value (second-worst)

Agents, sandboxes, & vibe-coded apps are powered by Cloud Run.

Sandbox Benchmarks

Q: What is Time to Interactive (TTI)?

TTI is the total elapsed time it takes for a sandbox provider to boot up a sandbox and run a terminal command inside the sandbox.

Q: Why use median instead of average?

The average can be easily skewed by a single extreme outlier. The median provides a more accurate representation of the typical performance most users will experience.

A leaderboard of common benchmarks for each of our sandbox providers.

View on Github

Last run: October 16, 2026

Want to see where you stand before benchmark Friday?Get stealth benchmarks →

Methodology

Performance Over Time

Composite Score

Detailed Metrics

Provider	Score	Median	P95	P99	Success
Isorun	99.5	0.05s	0.06s	0.06s	100%
Miosa	99.0	0.09s	0.10s	0.10s	100%
GravixlayerPre-release	98.1	0.19s	0.20s	0.20s	100%
Archil	96.7	0.20s	0.32s	0.86s	100%
CreateOS	97.6	0.23s	0.26s	0.26s	100%
sfcompute	95.5	0.39s	0.53s	0.54s	100%
Superserve	94.6	0.42s	0.70s	0.75s	100%
Vercel	94.6	0.49s	0.60s	0.65s	100%
Cloud Run	93.1	0.55s	0.89s	0.95s	100%
Blaxel	93.4	0.64s	0.68s	0.68s	100%
Arker	93.1	0.65s	0.73s	0.75s	100%
Daytona	91.9	0.72s	0.93s	0.96s	100%
Tenki	92.0	0.78s	0.83s	0.84s	100%
Cloudflare	87.5	0.83s	1.15s	1.30s	97%
Beam	90.2	0.93s	1.05s	1.05s	100%
Modal	87.3	0.94s	1.75s	1.78s	100%
Buddy	89.8	0.95s	1.10s	1.11s	100%
Mosaic	87.7	1.06s	1.41s	1.60s	100%
Declaw	87.0	1.19s	1.47s	1.48s	100%
E2B	86.3	1.25s	1.48s	1.68s	100%
Novita	76.5	1.71s	2.40s	4.78s	100%
Sandbox0	23.5	2.17s	28.53s	29.15s	50%
Sail	73.5	2.50s	2.69s	2.72s	99%
Tensorlake	74.3	2.53s	2.61s	2.62s	100%
Microsandbox	72.3	2.63s	2.91s	3.11s	100%
Upstash	22.5	6.65s	9.37s	9.43s	100%
Freestyle	0.0	13.15s	14.81s	14.85s	100%
Run Cloud	0.0	22.49s	81.01s	83.81s	69%
CodeSandbox	0.0	32.24s	38.71s	39.57s	88%
Hopx	0.0	0.00s	0.00s	0.00s	0%
Northflank	0.0	0.00s	0.00s	0.00s	0%

On this leaderboard? Your next public run is benchmark Friday — know your number every day.Get stealth benchmarks →

Want to see a provider added?

Let us know on X

Methodology

What We Measure

Every benchmark measures Time to Interactive (TTI) — the elapsed time from calling compute.sandbox.create() to the first successful runCommand() inside the sandbox.

Each provider is tested with 100 iterations per run. Benchmarks run automatically via GitHub Actions on a recurring schedule. All results are committed to the public benchmarks repo.

Burst Test: All sandboxes are launched concurrently in a single burst.

How We Score

The Composite Score is a weighted blend of timing metrics multiplied by the success rate. Each metric is scored against a fixed 10-second ceiling: 100 × (1 − value / 10,000ms), so a 200ms median scores 98 and anything ≥10s scores 0.

The weighted timing score is then multiplied by the success rate (0–1), so providers that fail frequently are penalized proportionally.

• Median: 60% — primary signal for typical experience
• P95: 25% — tail latency / consistency
• P99: 15% — extreme tail latency