# 100 computers at once: Workpaper on ComputeSDK's Sandbox Burst TTI

11 October 2026 · Workpaper engineering · https://workpaper.co/benchmark/sandbox-tti

## Key results

- Workpaper started 100 of 100 sandboxes requested at the same instant on ComputeSDK's Sandbox Burst TTI benchmark.
- Median time from request to first command: 952 ms. p95: 999 ms. p99: 1,003 ms.
- Score: 90.3 out of 100, which would place about 14th of 28 providers on ComputeSDK's board of 9 October 2026.
- On the server itself, the same burst of 100 took 0.36 s median.
- From India, a single Workpaper computer runs its first command in about 0.5 s.
- This was a self-run on 11 October 2026 on a single-node test deployment, not an official ComputeSDK listing.

## What the benchmark measures

ComputeSDK's Sandbox Burst TTI (https://www.computesdk.com/benchmarks/sandboxes) requests 100 sandboxes at once and times each from create() until its first command, node -v, succeeds. The fastest and slowest 5% are trimmed. Score = (0.6 × median + 0.25 × p95 + 0.15 × p99) × share started, where each term is 100 × (1 − t / 10 s).

## Where 90.3 would sit (board of 9 October 2026)

| Rank | Provider | Score |
|---|---|---|
| 9 | Daytona | 92.7 |
| 10 | Vercel | 92.6 |
| 11 | Cloud Run | 92.6 |
| 12 | Cloudflare | 92.2 |
| 13 | Tenki | 91.1 |
| – | Workpaper (self-run) | 90.3 |
| 14 | Beam | 90.2 |
| 15 | Modal | 88.5 |
| 16 | Declaw | 88.3 |
| 18 | Blaxel | 86.6 |
| 19 | E2B | 85.3 |

## How Workpaper starts 100 sandboxes at once

- Pre-booted pool: create requests claim an already-running computer close to the caller; a replacement boots behind it.
- Large memory pages: computers resume from snapshots backed by large pages, so a fresh computer's first node -v takes about 8 ms instead of about 300 ms. On the server, a burst of 100 went from 1.84 s to 0.36 s median.
- Admission queue: bursts wait in a fair queue instead of being rejected, so all 100 start.

## What limited this result

- One node: all 100 computers ran on a single machine.
- Distance: the runner was in Texas and the node in Europe, so every first command crossed the Atlantic.
- Test deployment: production capacity today is smaller and hosted in India.
- Self-run: ComputeSDK's official daily runs use their own runners; treat the placing as an estimate.

## Method

- Harness: https://github.com/computesdk/benchmarks (MIT), benchmarks/sandbox/tti.bench.ts, unmodified apart from a Workpaper provider entry.
- Command: --provider workpaper --iterations 100 --concurrency 100 (same flags as ComputeSDK's daily run).
- Runner: GitHub Actions, San Antonio, Texas, 11 October 2026.
- Raw data: https://github.com/ujj-exe/workpaper-benchmarks (all 100 timings, harness output, score script).

## Try Workpaper

Install: curl -fsSL https://workpaper.co/install | sh — 60 free minutes, no account needed. Pricing: ₹1 / $0.010 an hour for 2 vCPU, 4 GB; ₹2 / $0.020 an hour for 4 vCPU, 8 GB, 20 GB disk (--size large). Docs: https://docs.workpaper.co