English
Русский
简体中文
繁體中文
Tiếng Việt
Bahasa Indonesia
Português
EspañolAndroidWorld benchmark
How our agent scores on a cloud phone.
AndroidWorld is an open benchmark from Google Research: 116 tasks in 20 everyday Android apps, each checked by code against the state of the phone. We run it on a Clousd cloud phone with the same agent that does your tasks.
23 tasks that depend on today's date run on 25 Oct 2026; their results will be added here.
Results by app.
Each task gets one attempt. A task counts as passed only when the benchmark's own check accepts the final state of the phone.
| App | Tasks | Passed | Result |
|---|---|---|---|
| Simple Calendar Pro | 17 | 4 · 12 pending | 80% |
| Settings | 16 | 16 | 100% |
| Markor | 14 | 10 | 71% |
| Broccoli recipes | 13 | 6 | 46% |
| Pro Expense | 9 | 8 | 89% |
| Simple SMS Messenger | 6 | 6 | 100% |
| Tasks.org | 6 | 0 · 5 pending | 0% |
| OpenTracks | 6 | 0 · 6 pending | — |
| Joplin | 4 | 3 | 75% |
| Retro Music | 4 | 3 | 75% |
| Chrome | 3 | 2 | 67% |
| Clock | 3 | 3 | 100% |
| OsmAnd | 3 | 3 | 100% |
| Audio Recorder | 2 | 2 | 100% |
| Camera | 2 | 1 | 50% |
| Contacts | 2 | 2 | 100% |
| Files | 2 | 2 | 100% |
| VLC | 2 | 1 | 50% |
| Simple Gallery Pro | 1 | 1 | 100% |
| Simple Draw Pro | 1 | 1 | 100% |
How we ran it.
An ordinary Clousd phone
A Pixel 6a with Android 13, the same kind of cloud phone you create in the dashboard, with the benchmark apps installed by its setup.
The agent you use
The Clousd agent behind Tasks and POST /v1/runs, on the Clousd model: it reads the screen and its element tree, then taps, types and swipes until it decides the task is done.
The benchmark's own checks
Tasks, data and checks come from the AndroidWorld code unchanged. Only the connection differs: the harness reaches the phone through our adapter instead of a USB cable.
No second chances
One attempt per task, with the step limit the benchmark sets for its difficulty. Tasks that were run again after we fixed the phone or the adapter count their last run.
Where it still fails.
We read the steps of every failed task. None of them failed because of the phone; these are limits of the agent we are working on.
Long text entry: typing several full recipes or expenses copied from another app runs out of the step budget.
RecipeAddSingleRecipe, RecipeAddMultipleRecipesFromMarkor, RecipeAddMultipleRecipesFromMarkor2, RecipeAddMultipleRecipesFromImage, ExpenseAddMultipleFromMarkor, MarkorCreateNoteFromClipboard
Searching and counting in long lists: duplicates, items with a condition, totals over many rows.
RecipeDeleteDuplicateRecipes2, RecipeDeleteDuplicateRecipes3, RecipeDeleteMultipleRecipesWithConstraint, SimpleCalendarDeleteEvents, TasksHighPriorityTasks, NotesTodoItemCount, RetroPlaylistDuration
Watching and timing: reading text from a playing video, waiting for a recording, clicking a button and remembering changing numbers.
MarkorTranscribeVideo, CameraTakeVideo, BrowserMultiply
File chores across folders and apps: merging and moving notes, two playlists in one go.
MarkorMergeNotes, MarkorMoveNote, VlcCreateTwoPlaylists
Questions about the benchmark
What is AndroidWorld?
An open benchmark for agents that use Android phones, published by Google Research in 2024. It has 116 tasks in 20 apps such as Markor, Simple Calendar, OsmAnd, VLC and the system settings; every task is generated with fresh parameters and checked by code.
Why are some tasks not scored yet?
Tasks about "today" or "next week" were written for a fixed date in 2023. We moved their data to a Sunday in 2026 and run them on that day, so the phone's real date matches the task.
Can I run the benchmark on my own Clousd phones?
Yes. Create a phone, connect it to the AndroidWorld harness through the API, and drive it with your own agent over MCP or /v1/devices/{name}/observe and /act.