Phone-use agents

APK Arena

Levels
49
Models
21
Best
81.3%
gpt-6-astra
Field avg
58.5%
What it is

A suite of 49 individual phone-use levels across seven categories: Primitives, Vision, Memory, IQ, Tempospatial, Games, and Tasks. Each level is inspired by a real failure we encountered while running agents in production.

A gesture that kept missing, a game the agent played badly, a task it could not finish twice in a row. Turning those into levels gives us fast, repeatable readings on how models and harness changes hold up against the work we actually do.

How the agent plays it

Every run uses the nunu.ai harness: an agent loop wrapped around a code-execution sandbox that can capture screenshots and interact with the device. The agent takes screenshots, writes code to analyze what it sees, and then acts on the device through taps and swipes.

There is no structured state, no accessibility IDs and no text description of the screen at any point. Many games expose no accessibility tree at all, so the harness does not rely on one anywhere — which also means models are tested the way a person actually operates a phone, with no dev hooks to fall back on.

Levels49hand-built, all original
Categories7one per capability
Score range0–100%per level, per model
Screen accessPixelsno accessibility IDs
Level categories
8 levels

Primitives

Touchscreen control and fine motor accuracy — taps, swipes, complex gestures.

8 levels

Vision

Reading the screen: counting, matching, visual search.

5 levels

Memory

Spotting information that matters and recalling it across a long task.

7 levels

IQ

Reasoning and rule induction, mostly puzzles with no instructions.

8 levels

Tempospatial

Temporal and spatial reasoning under a clock.

7 levels

Games

Multi-step games that need a strategy, not just a correct answer.

6 levels

Tasks

Real workflows: driving phone UI, following multi-step instructions.

Scoring

Games

Scored on in-game performance over a fixed time. The 50% baseline is a human playing the same game for the same length of time.

Vision, memory, IQ, primitives

Straight accuracy. Did you count the right number of objects? Did you tap the right target? Black and white.

Tasks

Scored programmatically on how much of the workflow completed, with penalties for specific mistakes.

Every level returns a score between 0 and 100%. What that number means depends on the level, which is the point — one figure per level, comparable across models and across changes to our own harness. A level is scored on the agent's best attempt within the run.