Macaron: leaderboard
Metric: Overall accuracy (%), the unweighted mean of the multiple-choice and true/false accuracies in English and the local language, over the Macaron scenarios: 1,977 culturally grounded questions written by native annotators from 100 language-agnostic templates (7 reasoning types, 22 cultural aspects) for 20 cultural contexts and languages; zero-shot; open-weight models up to 8B scored by option log-probability, API models by a final-line JSON answer; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 84.7 | #145 |
| 2 | Gemini 3 Flash (Preview) (Non-reasoning) | 83.9 | #78 (Gemini 3 Flash (Preview)) |
| 3 | Claude Opus 4.5 | 78.4 | #79 |
| 4 | GPT-5 Chat | 77 | #233 |
| 5 | Gemini 2.5 Flash (Non-reasoning) | 75.7 | #237 (Gemini 2.5 Flash) |
| 6 | DeepSeek V3.1 (Thinking) | 75.1 | #260 (DeepSeek V3.1) |
| 7 | Gemini 2.5 Flash (Thinking) | 71.2 | #237 (Gemini 2.5 Flash) |
| 8 | Qwen 3 235B A22B 2507 Instruct | 68.7 | #291 |
| 9 | DeepSeek V3.1 (Non-reasoning) | 68.6 | #260 (DeepSeek V3.1) |
| 10 | Claude Haiku 4.5 | 66.9 | #271 |
| 11 | GPT-4o Mini | 66.8 | #588 |
| 12 | Llama 3.3 70B Instruct | 65.6 | #520 |
| 13 | Llama 4 Maverick | 65.6 | #451 |
| 14 | Qwen 2.5 7B Instruct | 52.3 | #846 |
| 15 | Llama 3.1 8B Instruct | 51.9 | #1018 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=macaron · How It Works · Data refreshed daily, snapshot 2026-10-11.