ScratchWorld - Overall (Primitive Mode): leaderboard
ScratchWorld (Zhang et al., 2026) gives multimodal agents 83 Scratch programming tasks in four categories (Create, Debug, Extend, Compute), each checked by unit tests run in the Scratch VM. These are primitive-mode results: the agent sees screenshots with a list of interface elements and must build programs through low-level GUI actions (click, drag-and-drop, typing), so precise block placement matters as much as program logic. The score is the percentage of all 83 tasks on which every unit test passes; higher is better.
Metric: Success Rate (%). Source: huggingface.co. Saturation forecast: Around March 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 13.25 | #163 |
| 2 | Claude Sonnet 4.5 | 8.43 | #244 |
| 3 | Gemini 2.5 Pro | 6.02 | #262 |
| 4 | Qwen 3 VL 32B Instruct | 2.41 | #582 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scratchworld-overall-primitive-mode · How It Works · Data refreshed daily, snapshot 2026-11-12.