Design Creativity Bench
Code · 2026-10-08
Kombai's 80 open-ended UI design tasks, each written for two product domains (160 prompts), with one standalone HTML design per prompt. The overall score gives equal weight to standardised diversity (0.9 originality + 0.1 creative range) and standardised appropriateness, reported as a T-score over the paper's frozen 15-model panel: 50 is the mean panel model and 10 one panel standard deviation, so it is unbounded and panel-relative. It is a diagnostic because the panel-relative T-score has no fixed bounds.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 65.5 |
| GPT-6.1 Sol | 63.7 |
| Grok 4.7 | 62.2 |
| Gemini 3.8 Flash | 60.6 |
| GPT-5.6 Sol | 51.3 |
| Claude Opus 5.5 | 51.0 |
| Claude Fable 5.1 | 48.6 |
| Muse Spark 1.3 | 47.2 |
| DeepSeek V4.1 Flash | 43.6 |
| Grok 4.6 | 43.4 |
| GLM-5.3 | 41.9 |
| Kimi K3 | 40.4 |
| GLM-5.3 Flash | 38.5 |
| MiniMax M3 | 34.0 |