Benchmarks

Every number here is measured and dated. The caveats stay with the comparisons.

Session create → first command

The ComputeSDK method, run on 2026-07-16: wall-clock sandbox.create() plus the first successful runCommand('node -v'), 5% trimmed. The composite weighs normalized scores of the median (0.60), P95 (0.25), and P99 (0.15) on a 10-second scale, times the success rate.

Mode	N	Success	Median	P95	P99	Composite
Sequential	100	100/100	0.61 s	0.82 s	0.87 s	93.0
Staggered (200 ms)	100	100/100	0.57 s	0.78 s	0.92 s	93.3
Burst (simultaneous)	100	100/100	1.15 s	1.68 s	1.96 s	86.0

These are per-request Durable Object boots with no warm pool. All three runs had 100% success. In a phase run (N=15), create()+ready() took a 545 ms median and first exec took 23 ms. Session boot is ~96% of TTI; after that, exec round-trips take ~20 ms. Across ~370 total sandboxes there was one failure: a DO storage timeout reset during a smoke iteration.

The client was a residential Linux machine, so network RTT to the edge is in these numbers. Nimbus is not on the official ComputeSDK leaderboard; the unmodified harness ran locally. Treat cross-provider numbers as directional. In the same-platform comparison, Nimbus measured ~3× faster across the board than Cloudflare’s container-based Sandbox SDK using its published leaderboard data from 2026-07-13: 0.61 s vs 1.95 s sequential median, and 1.15 s vs 4.05 s burst.

Everyday operations

Measured against the live deployment in 2026-07.

Operation	Wall time
git clone facebook/react (7,300 files)	~28 s
npm install 611-package app (52,500 files, cold)	~165 s
npm install zod (cold session)	~6 s
node -e 'console.log(...)' (warm)	102–152 ms
Vite hot reload	302 ms median

Cross-session caching gives 9–16× warm-vs-cold speedups on package installs.

tests/behavioral/perf-regression/ checks these thresholds on every deploy. Anyone can run it:

BASE=https://nimbus-os.dev bun test:behavioral