CreativityBench: leaderboard
CreativityBench (2026) tests creative tool use: given a household scene listed as entities with their parts and physical and state attributes, the model must find the part of an unconventional object whose affordance accomplishes a task, among distractors. Score is the percent of the 14,000 generated tasks where both the chosen entity and part match the gold; higher is better.
Metric: Gold correct rate (%, times 100): share of tasks where the model selects both the gold entity and the gold part, exact match, over the 14K affordance-based creative tool-use tasks of CreativityBench (eight household scenes; the model sees the task and every entity in the scene with its parts and attributes, and must name the entity and part to repurpose and how to use it), temperature 0 except GPT-5 Mini and Nano (fixed default sampling); higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 10 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 32B | 25.88 | #972 |
| 2 | Qwen 3 14B | 24.83 | #1269 |
| 3 | Llama 3.3 70B Instruct | 21.51 | #1385 |
| 4 | Ministral-3-14B-Reasoning-2512 | 20.91 | #1457 |
| 5 | Qwen 3 4B | 18.82 | #2348 |
| 6 | GPT-5.2 | 18.19 | #185 |
| 7 | GPT-5 Mini | 16.87 | #364 |
| 8 | Gemini 2.5 Pro | 16.7 | #276 |
| 9 | Gemini 2.5 Flash | 15.32 | #463 |
| 10 | GPT-5 Nano | 11.92 | #906 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=creativitybench · How It Works · Data refreshed daily, snapshot 2026-11-21.