Sandboxes and code execution

Sandboxes are judged on isolation and unit cost, not intelligence. The 2026 benchmark activity is about escapes: the UK AI Security Institute and Oxford published SandboxEscapeBench in March 2026. A useful ranking here would grade isolation and escape audits, which nobody publishes yet.

1 independent chart · rest catalog Category 8 of 11
Independent benchmarkaiworkflowlab controlled run (third party) · read Nov 7, 2026

Cold start

  1. Daytona 90 ms
  2. E2B 148 ms
  3. Cloudflare 213 ms
  4. Modal 407 ms

Shorter is faster. Same run, 1M 30-second sessions at 1 vCPU: Cloudflare ~$14, Daytona ~$18, Modal ~$20.

Tools

Alphabetical. Links go to each vendor's own site.

Catalog only, not yet measuredvendor pages and secondary sources · compiled Nov 7, 2026
ToolWhat it doesEvidencePrice or terms
Cloudflare Sandbox Workers lineage; generous free tier 213 ms cold start ~$14 per 1M short sessions
Daytona containers + VM class; open-source repo archived Jul 2026 90 ms cold start $0.0504/vCPU-h; $200 free
E2B Firecracker microVMs; open-source infra 148 ms cold start free $100 credits; Pro $150/mo + usage
Microsandbox libkrun microVMs; self-host alternative none yet free, self-hosted
Modal Sandboxes gVisor; GPUs everywhere 407 ms cold start $0.1419/core-hr; $30/mo free
Runloop VM-backed devboxes none yet usage-based; $50 credits

Caveats

  • Escape benchmarks (SandboxEscapeBench, AgentEscapeBench) test agents against sandboxes; they do not publish vendor-by-vendor results.

Other categories