README.md

---
license: cc-by-4.0
task_categories:
- image-to-text
- visual-question-answering
language:
- en
- de
- fr
- es
- it
- nl
tags:
- gui-grounding
- computer-use
- benchmark
- screenshots
- synthetic-data
- spreadsheets
- text-grounding
- professional-apps
pretty_name: Pointerbench
size_categories:
- 1K<n<10K
---

# Pointerbench

Pointerbench is a small GUI grounding benchmark suite for computer-use models.
Each example has one screenshot, one instruction, target geometry in absolute
pixels, and a binary evaluation rule.

Links:

- GitHub: https://github.com/warmwindOS/pointerbench
- Blog post: https://about.warmwind.com/pointer-bench/
- Add your model to the official benchmark leaderboard: https://warmwind.com/contact

The suite has three subsets:

| Subset | Examples | What it tests |
| --- | ---: | --- |
| `pointerbench-sheets` | 500 | Spreadsheet cells, colors, headers, edges, corners, and relative positions |
| `pointerbench-text` | 500 | Words, characters, punctuation, caret positions, chrome text, text bounding boxes, and invoice fields |
| `pointerbench-pro` | 500 | Icons, text, and mixed GUI targets across 100 professional applications |

## Leaderboard

Scores are percentages. The table is ranked by the reported public average across Sheets, Text, and Pro. The Fable 5 row is a uniform 50 percent sample rerun: 250 examples per benchmark, seed 20260702. Other rows are full public report runs unless otherwise noted.

| Rank | Model | Sheets | Text | Pro | Average |
| ---: | --- | ---: | ---: | ---: | ---: |
| 1 | `anthropic/claude-fable-5` | 96.0 | 58.4 | 80.8 | 78.7 |
| 2 | `warmwind/pointer-1.6 (agentic)` | 87.8 | 66.8 | 76.8 | 77.1 |
| 3 | `anthropic/claude-sonnet-4.6` | 82.6 | 47.8 | 77.4 | 69.3 |
| 4 | `anthropic/claude-opus-4.8` | 82.2 | 44.4 | 80.2 | 68.9 |
| 5 | `warmwind/pointer-1.6` | 77.0 | 49.2 | 76.8 | 67.7 |
| 6 | `anthropic/claude-opus-4.7` | 73.8 | 36.2 | 75.4 | 61.8 |
| 7 | `openai/gpt-5.5` | 72.4 | 31.4 | 75.8 | 59.9 |
| 8 | `openai/gpt-5.4` | 63.2 | 25.0 | 50.4 | 46.2 |
| 9 | `openai/gpt-5` | 35.2 | 13.6 | 20.2 | 23.0 |
| 10 | `x-ai/grok-4.20-multi-agent` | 10.6 | 1.4 | 10.0 | 7.3 |
| 11 | `moonshotai/kimi-k2.5` | 10.8 | 3.4 | 3.0 | 5.7 |
| 12 | `openai/gpt-5-mini` | 6.0 | 1.4 | 5.8 | 4.4 |
| 13 | `moonshotai/kimi-k2.6` | 5.6 | 2.8 | 2.8 | 3.7 |
| 14 | `qwen/qwen3.7-plus` | 5.8 | 0.6 | 3.0 | 3.1 |
| 15 | `qwen/qwen3.6-flash` | 4.2 | 1.2 | 2.6 | 2.7 |
| 16 | `minimax/minimax-m3` | 3.6 | 0.4 | 3.6 | 2.5 |
| 17 | `qwen/qwen3.6-plus` | 3.0 | 0.6 | 3.6 | 2.4 |
| 18 | `qwen/qwen3-vl-235b-a22b-thinking` | 2.6 | 0.4 | 3.2 | 2.1 |
| 19 | `qwen/qwen3.5-9b` | 2.0 | 0.2 | 2.6 | 1.6 |
| 20 | `google/gemini-3.5-flash` | 3.4 | 0.6 | 0.6 | 1.5 |
| 21 | `openai/gpt-5-nano` | 2.0 | 0.2 | 1.6 | 1.3 |
| 22 | `moonshotai/kimi-k2.7-code` | 0.4 | 1.8 | 0.8 | 1.0 |
| 23 | `x-ai/grok-4.3` | 1.2 | 0.0 | 0.0 | 0.4 |
| 24 | `google/gemini-3.1-pro-preview` | 0.0 | 0.2 | 0.4 | 0.2 |
| 25 | `google/gemini-3.1-pro-preview-customtools` | 0.0 | 0.4 | 0.0 | 0.1 |
| 26 | `stepfun/step-3.7-flash` | 0.0 | 0.0 | 0.2 | 0.1 |
| 27 | `x-ai/grok-build-0.1` | 0.0 | 0.0 | 0.0 | 0.0 |
| 28 | `xiaomi/mimo-v2.5` | 0.0 | 0.0 | 0.0 | 0.0 |
| - | Center-click baseline | 0.4 | 0.2 | 0.2 | 0.3 |

To add your model to the official benchmark leaderboard, contact https://warmwind.com/contact.

All images are synthetic 1024x768 PNG screenshots. The datasets contain no
scraped user data and no PII.

## Layout

Each subset is self-contained:

```text
pointerbench-sheets/
  data/test/metadata.jsonl
  data/test/0000.png
  eval.py
  README.md
  REPRODUCE.md

pointerbench-text/
  data/test/metadata.jsonl
  data/test/0000.png
  eval.py
  README.md
  REPRODUCE.md

pointerbench-pro/
  data/test/metadata.jsonl
  data/test/0000.png
  eval.py
  README.md
  REPRODUCE.md
```

## Schema

Each metadata row includes:

```json
{
  "file_name": "0000.png",
  "id": "pbs_0000",
  "instruction": "Click cell E11.",
  "bbox": [x1, y1, x2, y2],
  "point": [x, y],
  "answer_type": "point",
  "eval": {"type": "point_in_bbox", "bbox": [x1, y1, x2, y2]},
  "data_type": "cell",
  "category": "cell_ref",
  "image_size": [1024, 768]
}
```

Point tasks are correct when the predicted point lands inside the target bbox.
Bbox tasks, used in Pointerbench-Text, use an asymmetric overlap rule: a hit
requires the ground truth to be almost fully covered (coverage >= 0.90) and the
prediction to stay reasonably tight around it (precision >= 0.70). This
penalises predictions that cut off part of the target far more than predictions
that wrap it with some margin.

## Evaluation

Run the scorer inside a subset folder:

```bash
python eval.py --predictions preds.jsonl
```

Predictions are JSONL rows with an `id` and either a `point` or `bbox`, depending
on `answer_type`.

Recommended inference prompt:

```bash
python eval.py --show-system-prompt
```

```text
You are evaluating Pointerbench, a GUI grounding benchmark. You will receive one 1024x768 screenshot and one task instruction. Use absolute pixel coordinates with origin at the top-left of the image. Do not return normalized coordinates. Do not crop or resize the coordinate frame. For point tasks, return JSON like {"point": [x, y]}. For bounding-box tasks, return JSON like {"bbox": [x0, y0, x1, y1]}.
```

You can edit the prompt for your inference stack. Keep the 1024x768 absolute
pixel coordinate frame fixed, and report any image resizing or multi-step zoom
strategy with your results.

See each subset README for the exact distribution, schema details, and examples.

## License

Dataset images and annotations are released under CC BY 4.0. The included
evaluation scripts are released under MIT.