AI sandbox benchmark
Last updated 2026-07-26 · figures are vendor-published; corrections welcome
| Sandbox | Isolation | Cold start | $/vCPU-hr | GPU | Session cap |
|---|---|---|---|---|---|
| Blaxel | microVM | 25ms | $0.083 | No | Persistent (hibernation) |
| Cloudflare Sandbox SDK | V8 isolate + container | 50ms | $0.072 | No | 30min execution cap |
| Daytona | Docker container | 90ms | $0.05 | Yes | Persistent (lifecycle-managed) |
| E2B | Firecracker microVM | 150ms | $0.05 | No | Up to 24h |
| Northflank | Kata / gVisor microVM (selectable) | 300ms | $0.01667 | Yes | Unlimited |
| Fly.io Machines | microVM | 300ms | $0.07 | Yes | Persistent |
| Vercel Sandbox | Firecracker microVM | 400ms | $0.128 | No | 45min (Hobby) / 24h (Pro) |
| Runloop | microVM | 400ms | $0.108 | No | Configurable |
| CodeSandbox SDK | Firecracker microVM | 500ms | $0.05 | No | Configurable |
| Modal | gVisor | 800ms | $0.14 | Yes | Configurable |
Method: values are the providers' own published cold-start and pricing figures as of 2026-07-26, normalized to milliseconds and USD/vCPU-hour. Cold start varies with image size, region and warm-pool state — treat these as directional. Independent measured runs are on the roadmap; this table will carry them when live.
How to read this
Cold start and headline price are the two numbers buyers anchor on, but they rarely decide the bill alone. Isolation strength (microVM vs container), idle/paused cost, session caps, GPU availability and creation fees all move real spend. Use this table to shortlist, then model your actual workload in the cost calculator.
FAQ
Which AI sandbox has the fastest cold start?
As of 2026-07-26, Blaxel has the fastest published cold start in our benchmark at 25ms, using microVM.
Which AI sandbox is cheapest per vCPU-hour?
Northflank has the lowest published rate at $0.01667 per vCPU-hour, though total cost depends on idle time and session length.
Every sandbox worth knowing, once a month.
The refreshed benchmark, pricing changes, new entrants and one sharp take. Read by engineers choosing where their agents and apps run. No fluff.