blolabel/phoneshell-bench

PhoneShell: an iOS agent benchmark measured on a physical iPhone

80 task definitions with machine-checkable success conditions, and 397 scored runs of seven frontier models against them. Every run happened on a physical iPhone over a USB cable. No simulator, no emulator, no mock. A task passes only when the phone itself ends in the required state, so what a model says it did counts for nothing.

Harness: github.com/Blomega/phoneshell · Results: blolabel.ai

---

The finding this dataset exists to support

Vision is worth 100 points or zero, and nothing in between.

Ten near-identical web pages were built, each carrying one word written into the HTML and a different word rendered into an image with no alt text. On the device, 0 of 10 image words appear anywhere in the accessibility tree and 10 of 10 written words do. Three models were then asked for each word, twice, once with a screenshot and once without, back to back with the order alternating.

where the answer lives	with the image	tree only	McNemar exact
only in the pixels	30/30 (100%)	0/30 (0%)	p < 0.0001
in the accessibility tree	29/30 (97%)	30/30 (100%)	p = 1.0

The interaction is the result, not either row alone. Same pages, same navigation, same question shape, same prompt, same resolution: the only thing that changed is where the answer sat. The control row shows nothing, so the effect is not an artifact of the harness, and carrying the screenshot costs 35% to 43% more per task with no change in step count.

The design was pre-registered before the run (EXPERIMENT-2.md), which matters here because the first version of this experiment reported the opposite and was wrong: the image had been silently downscaled to 235x512, so the vision arm was never a test of vision. That retraction is in the repository too.

The three models return identical numbers because the stimulus is deliberately binary. That is the predicted result and a replication, not a tie. It is the opposite situation to the one below, and telling the two apart is the whole point of reading the design before the numbers.

---

Read the leaderboard's own warning first

results/xv.jsonl holds six models over the same 24 tasks. Sorted by pass rate it looks like a ranking. It is not one.

The models ran the same tasks, so this is a paired design and the test is McNemar's exact on the disagreements. Zero of the fifteen model pairs reach p < 0.05. The widest gap on the board is 4 wins to 0 at p = 0.12 and would need roughly 33 tasks before it counted. Three pairs disagree perfectly symmetrically, 2 wins each, so no size of task set will ever separate them: to this instrument they are the same model.

12 of the 24 tasks were passed by every model and 2 by none. Both kinds carry zero information about ranking, so 58% of that run was measuring nothing and the real instrument was 10 items wide.

This is published as-is, with the arithmetic, because a benchmark that quietly presents six percentages as an order is the more common failure and the harder one to notice. Compute it yourself from these files.

---

Files

file	rows	what it is
tasks.jsonl	80	task definitions with their success conditions
results/xv.jsonl	144	cross-vendor sweep, 6 models x 24 tasks
results/xv-text.jsonl	48	the two text-only models on the same tasks
results/xv-novision.jsonl	24	GPT-5.1 again with the screenshot withheld
results/exp2-v.jsonl	60	experiment 2, image arm
results/exp2-n.jsonl	60	experiment 2, tree-only arm
results/v4.jsonl	66	single-model reference run over the whole suite (5 skipped)

Models covered: GPT-5.1, Gemini 3.1 Pro, Claude Opus 4.8, Kimi K3, Qwen3-Max, DeepSeek v3.2, Claude Sonnet 5. Qwen3-Max and DeepSeek v3.2 take no image and drive the phone from the accessibility tree alone.

A task

{
  "task_id": "clock.alarm.set_time",
  "instruction": "Open the Clock app and set an alarm for 6:45 AM.",
  "app": "com.apple.mobiletimer",
  "difficulty": "hard",
  "tags": ["capability:picker-set"],
  "checks": [{"kind": "element_exists", "text": "6:45", "negate": false}],
  "max_steps": 25
}

checks is the grader. element_exists and regex_on_screen read the device's accessibility tree after the agent stops; foreground_app reads the frontmost bundle id; answer_contains is the one check that grades what the model said, used only where the answer is a value the task itself planted.

A result

{
  "task_id": "probe.image.eight", "model": "openai/gpt-5.1", "passed": true,
  "turns": 3, "seconds": 23.97, "cost_usd": 0.003919,
  "checks": [{"kind": "foreground_app", "passed": true, "detail": "..."}],
  "agent_said": "The word shown in the picture on the page is NIMBUS."
}

skipped: true means the model was never asked, because the rig was unhealthy or the app was absent. Skipped rows are not failures and must be excluded, not counted as zero. Reporting them as failures blames the model for the device.

---

Loading

from datasets import load_dataset

tasks   = load_dataset("blolabel/phoneshell-bench", "tasks",   split="train")
results = load_dataset("blolabel/phoneshell-bench", "results", split="train")

scored = [r for r in results if not r["skipped"]]

Reproducing the headline test needs nothing but these files:

import math, collections

def mcnemar_exact(b, c):                       # two-sided exact, paired binary
    n = b + c
    if n == 0: return 1.0
    k = min(b, c)
    return min(1.0, 2 * sum(math.comb(n, i) for i in range(k + 1)) / 2 ** n)

b and c are the disagreements: tasks passed by one arm only, and by the other only. Tasks both arms passed carry no information about which is better, which is why the sample size never appears.

---

Scope, and what this does not measure

• Apple's own applications only. Settings, Safari, Notes, Reminders, Clock, Calculator, Contacts, Maps, Calendar, Files, Books, Compass, Shortcuts, Voice Memos and the home screen. No third-party app is automated, so no third party's terms are involved.
• First-party iOS is close to fully described by its accessibility tree. Measured across these screens, the tree covers most readable text, which is why the vision effect above had to be constructed rather than found. A chat app on the same phone returns WAMessageBubbleTableViewCell as a button name. Whether vision rescues a badly labelled third-party app is the commercially interesting question and is not answered here.
• One device. iPhone 17 Pro Max on iOS 26. A different phone is a different instrument.
• 16 tasks are held out and appear in neither this dataset nor the repository nor its history. A benchmark whose entire answer key is public becomes training data, and the score then measures memorisation.
• No personal data. The runs happened on a personal phone. Free-text model output is kept only where a grader checks it against a value the task planted, and the device identifier is stripped. Everything else the model said is dropped rather than reviewed.

Citation

@misc{phoneshell2026,
  title  = {PhoneShell: measuring iOS agent capability on a physical device},
  author = {Shamsan, Sam},
  year   = {2026},
  url    = {https://github.com/Blomega/phoneshell}
}

Apache-2.0. The harness vendors WebDriverAgent (BSD-3-Clause).