Sandbox Arena
A benchmark that races two cloud code sandboxes on the same task.
InfraData
Open sandboxarena.vercel.appSource
The numbers
- 24commits
git rev-list --count 0dae422af - 1GitHub star
gh api repos/Kevin-Liu-01/Sandbox-Arena --jq .stargazers_count (2026-11-01) - 3check files
gh api repos/Kevin-Liu-01/Sandbox-Arena/git/trees/0dae422af4e9189d982e981746d2379f73a28d2e?recursive=1 --jq ".tree[].path" | grep -cE <glob> (tree not truncated, 220 entries) - 4commits per active day
24 commits over 6 distinct commit dates, 2026-06-14 to 2026-07-27
Trait
Fresh Sandbox Each run gets a fresh sandbox, deleted after, so its conditions clear every turn.
Sandbox-Arena@0dae422af:lib/arena/orchestrator.ts:37-38, 49-51, 78-80; components/arena/TrialStudio.tsx:266Moves
Side By Side hits for 40 twice, 95% of the time. Two sandboxes start the same task at once and stream their logs side by side, so you watch the race live.
Sandbox-Arena@0dae422af:lib/arena/orchestrator.ts:91-93, 121-128; README.md:7-10Verified Finish hits for 80 and finishes a target under 35% health. Every task ends by checking its output against a known answer, so a fast but broken run scores zero.
Sandbox-Arena@0dae422af:lib/arena/scoring.ts:100-101, 117-121; lib/arena/workloads.ts:30, 57, 99, 137, 169, 206, 247, 272, 306 (verifies); app/methodology/page.tsx:88-89Slowest Runs hits for 95, 90% of the time. Rankings can sort providers by the time that 1 run in 20 exceeds, because averages hide rare long stalls.
Sandbox-Arena@0dae422af:components/benchmarks/BenchmarkExplorer.tsx:13, 40-44, 93-94; app/benchmarks/page.tsx:45-47; README.md:26-28Session Keys raises a 40% barrier. Your own account keys run one real trial, are read once, expire in ten minutes and never reach logs or the sandbox.
Sandbox-Arena@0dae422af:lib/arena/credentials.ts:5-6, 63-70; lib/arena/security.ts:113-120; README.md:36-38
Numbers pinned at 0dae422af, 2026-07-27.



