Reliability Benchmarks

Our reliability benchmark creates a sandbox, runs a workload, and destroys it every minute to test the uptime reliability of sandbox providers. All times are UTC.

All passedBlipsDownMissed (scheduler)Pending
ParticipantUptimeDowntimep50
Schedulerours99.41%3 missed—
Promigence100%—146ms
Miosa100%—186ms
Archil100%—267ms
CreateOS100%—287ms
Daytona100%—400ms
Superserve100%—435ms
Blaxel100%—480ms
Beam100%—508ms
Novita100%—816ms
Mosaic100%—867ms
Upstash100%—895ms
Vercel100%—1.74s
Namespace100%—2.19s
Tensorlake100%—2.24s
Run Cloud100%—4.36s
Isorun99.80%1 min229ms
Gravixlayer99.80%1 min252ms
Arker99.80%1 min410ms
Buddy99.80%1 min452ms
Cloudflare99.80%1 min670ms
Cloud Run99.80%1 min723ms
Modal99.80%1 min760ms
sfcompute99.80%1 min1.07s
Freestyle99.80%1 min1.24s
E2B99.80%1 min1.75s
Sandbox099.80%1 min3.02s
CodeSandbox99.80%1 min8.25s
Sprites99.60%2 min855ms
Railway99.60%2 min5.03s
Sail99.41%3 min3.23s
Tenki98.82%6 min977ms

Hover a cell for its numbers. Click a participant cell to open their checks for that day.

  • DownTenki
    Wed, Nov 4 08:28 UTC → Wed, Nov 4 08:31 UTC · 4 min · 3 failed
    step_timeout
  • DownSail
    Wed, Nov 4 04:51 UTC → Wed, Nov 4 04:54 UTC · 4 min · 3 failed
    step_timeout
What went wrong (4 errors)
  • 12×step_timeout
  • 10×participant failed
  • 2×probe_failed
  • 1×participant failed before reporting results

How it works: a ComputeSDK Actions run fires on schedule and, for each provider, creates a fresh sandbox through that provider’s ComputeSDK adapter — the same create path our TTI and DAX benchmarks use. Inside the sandbox it writes a 1 MiB file of random data, SHA-256 hashes it twice, requires the digests to match, deletes the file, and destroys the sandbox. The check passes only if the full create → probe → destroy loop succeeds in time; runs that fail to dispatch count as scheduler misses, not provider downtime.

Need a provider’s uptime on demand? Sign up for a free ComputeSDK account and query any provider’s status through the API, CLI, or MCP.

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