Butter-Bench: Infer Butter Bag
Agents · 2025-10-23
Butter-Bench subtask measuring whether a model can visually infer which package contains butter from refrigeration cues and package markings.
Top models (higher is better)
| Model | Score |
|---|---|
| Human (Butter-Bench) | 100.0 |
| Grok 4 | 80.0 |
| GPT-5 | 60.0 |
| Gemini 2.5 Pro | 40.0 |
| Gemini Robotics-ER 1.5 | 40.0 |
| Llama 4 Maverick | 40.0 |
| Opus 4.1 | 0.0 |