CulturalMenuBench - Multiple Choice: leaderboard
CulturalMenuBench (arXiv 2609.03526v1): culinary multimodal reasoning over 587 dishes from 18 regions with dish and step-by-step cooking images, ingredients and procedure text, prompted in each cuisine's language. Table 5 knowledge probe: four-way multiple choice over the same pairings for 12 models; accuracy percent, chance 25, higher is better.
Metric: Accuracy (%; four-way, chance 25). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 96.8 | #105 |
| 2 | GPT-5.1 (2025-11-13) | 96.6 | #107 |
| 3 | Claude Opus 4.6 | 96.2 | #67 |
| 4 | Qwen 3.6 Plus | 95.1 | #124 |
| 5 | Kimi K2.5 | 94.7 | #152 |
| 6 | GPT-5 Mini (2025-08-07) | 94.7 | #175 |
| 7 | Gemini 2.5 Pro | 94.6 | #159 |
| 8 | Gemma 4 26B A4B (IT) | 88.4 | #293 |
| 9 | Claude Sonnet 4 | 87.1 | #207 |
| 10 | GPT-5 Nano | 82.2 | #439 |
| 11 | Qwen 3.5 35B A3B | 78.8 | #268 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=culturalmenubench-multiple-choice · How It Works · Data refreshed daily, snapshot 2026-11-05.