AndroidWorld benchmark
mobile-use solved all 116 AndroidWorld tasks.
Our open-source agent, on Google DeepMind’s benchmark. mini is the QA agent we built next.
mobile-use on AndroidWorld
116/116
- mobile-useopen source
- google/artemisuses mobile-use
miniour QA agent, built next
How to read this result.
What was measured
Tasks in plain words on an Android emulator, checked against device state.
How to check it
The code is open source, the paper is public, every task is below.
When
116 of 116 in January 2026. The paper followed on 8 February.
What it doesn’t tell you
QA accuracy on production apps. mini has no score here.
All 116 tasks, by app.
By AndroidWorld id. Open one to see mobile-use’s recorded run.
Simple Calendar Pro17/17 solved
Settings16/16 solved
Markor14/14 solved
Broccoli13/13 solved
Pro Expense9/9 solved
OpenTracks6/6 solved
Simple SMS Messenger6/6 solved
Tasks6/6 solved
Joplin4/4 solved
Retro Music4/4 solved
Chrome3/3 solved
Clock3/3 solved
OsmAnd3/3 solved
Audio Recorder2/2 solved
Camera2/2 solved
Contacts2/2 solved
Files2/2 solved
VLC2/2 solved
Simple Draw Pro1/1 solved
Simple Gallery Pro1/1 solved
Good questions.
Did mini score 116 of 116?
No. mobile-use, our open-source agent, did. mini is the QA agent we built next. It writes, runs and maintains tests on iOS, Android and web, and has no AndroidWorld score.
Why does the paper say Minitap?
The paper names the system after the company. Its code is mobile-use, linked from the abstract. mini is a different, newer agent.Read the paper
How was it measured?
Our paper reports the result under the official AndroidWorld evaluation protocol, without modification, and lists the models used. Google’s Artemis uses mobile-use.
How do other agents compare?
In our 8 February 2026 paper (Table 1): AGI-0 97.4%, askui 94.8%, gbox.ai 86.2% and Z.AI 80.2%. H Company reported 87.1% for Surfer 2 in October 2025. These are published results, not a live leaderboard.See Table 1
What does this mean for testing my app?
AndroidWorld shows an agent can finish real tasks on an Android emulator. Testing your app needs more. mini learns your product, writes the tests and their acceptance criteria, runs them on real devices, simulators and browsers, and keeps the suite current as your app changes.How mini tests your app
Bring your hardest flow.
Bring a build. We run mini on it during the call, and you leave with the report.