On November 8, 2026, Limrun gave our script a usable cloud Android phone in under a second (0.98 s) at the median, Mobilebox in 19.4 seconds and Phone Harness in 23 seconds. Mobilebox, which runs cloud Android phones for AI agents at $0.01 a minute, was the cheapest for a 10-minute session ($0.10) and the only one where we measured park and resume; Limrun was also fastest from tap to screenshot (0.38 s, against 1.45 s for Mobilebox and 1.9 s for Phone Harness).
These are our own measurements: 10 sessions per service from one client in Nuremberg, Germany, on each vendor’s free tier or trial, with the same steps and the same APK. The benchmark page has every metric, the p90s, the CSV and the scripts.
The numbers
| Service | Cold start to first screenshot | Tap to screenshot | Install a 4 MB APK | 10-minute session |
|---|---|---|---|---|
| Limrun | 0.98 s (p90 1.07) | 0.38 s (p90 0.40) | 0.85 s | $0.40 on Entry |
| Mobilebox | 19.4 s (p90 19.5) | 1.45 s (p90 1.49) | 1.2 s | $0.10 |
| Phone Harness | 23.0 s (p90 70.0) | 1.9 s (p90 2.4) | 3.6 s, uploaded over ADB | $0.50 |
Medians, p90 by nearest rank, nothing removed. Prices are list prices on the day: Limrun $0.04 a minute on its Entry plan (pricing), Mobilebox $0.01 (pricing), Phone Harness $0.05 (limits and credits).
Why the starts differ so much
Limrun’s create call returned a ready phone in about a third of a second, so its phones were very likely booted before we asked. Mobilebox boots a phone per request, and Phone Harness’s timings look the same: Mobilebox in a narrow 18.9 to 19.5 seconds across all 10 runs, Phone Harness in two clusters, about 21 seconds and about 67 seconds. Phone Harness’s own docs say "about 80 seconds", so its faster runs beat its published number.
For an agent the difference is real. A one-minute task spends a third of its wall clock waiting on a 20-second start, and almost none on a one-second start. For a 30-minute session the start barely matters, and the price per minute and what a stopped phone keeps matter more.
Tap latency includes the network
We timed one tap plus one full screenshot through each vendor’s own interface: Limrun’s SDK, the Mobilebox REST API, and ADB for Phone Harness, which has no tap endpoint. Network distance counts twice. Our client was 8 ms from Mobilebox’s API, about 50 ms from Limrun’s and about 170 ms from Phone Harness’s servers at US West addresses. Phone Harness’s 252 ms taps are mostly that round trip; Mobilebox’s 517 ms taps, from 8 ms away, are not.
What a stopped phone keeps
Mobilebox parked in about 1 second and came back to a screenshot in 11.5 seconds, with the installed APK still there in 10 of 10 runs. Phone Harness documents a saved profile phone that "comes back the same way next time"; we didn’t test it. Limrun’s Android instances end when they are deleted or time out, so there is nothing to resume.
What we left out
BrowserStack, Genymotion and LambdaTest forbid publishing performance results or competitor use in their terms, so we didn’t measure them. Mobilerun, AWS Device Farm, Firebase Test Lab and phonebox.dev need an OAuth or Google sign-in or a payment card, which this run didn’t use. The benchmark page lists each reason and their published prices.
Run it yourself
The scripts are in fifteenA/mobilebox/benchmarks, one per service. Each prints one line per measurement, appends it to a CSV and deletes what it created. Results will differ from your location and on another day; that is why the client, date and raw samples are published.