DragOn: leaderboard
DragOn (2026) tests drag grounding for GUI agents: from one screenshot and an instruction the model outputs the start and end points of a drag, and both must land inside the pixel-aligned source and target regions. The score is the success rate on the private 1,000-example test split, 250 examples in each of four domains (text highlighting, spreadsheet cell selection, slider manipulation and slide element resizing). Higher is better.
Metric: Success rate (%) on the four drag domains (text highlighting, spreadsheet cell selection, slider manipulation and slide element resizing) of the private 1,000-example DragOn test split (250 per domain): from one screenshot and instruction the model outputs the drag's start and end points as JSON, and both must fall inside the pixel-aligned source and target boxes; each model at its documented settings (Claude extended thinking off, Kimi-K2.5 reasoning off, local Qwen and Holo models with chain-of-thought on); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 27.7 | #79 |
| 2 | GPT-5.4 | 25.7 | #123 |
| 3 | Qwen 3.5 397B A17B | 21.5 | #246 |
| 4 | Kimi K2.5 (Non-reasoning) | 11.6 | #256 (Kimi K2.5) |
| 5 | Claude Sonnet 4.5 | 10.7 | #250 |
| 6 | Qwen 3.5 35B A3B | 2.3 | #482 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dragon · How It Works · Data refreshed daily, snapshot 2026-11-13.