MAGI-Hard: leaderboard
A hard subset distilled from MMLU and AGIEval, kept because it discriminates between strong models where the parent benchmarks have flattened. Frozen 2024-05 snapshot with a roster to match — GPT-4o-0513 leads it.
Metric: Accuracy (%, 2024-05 snapshot). Source: huggingface.co. Saturation forecast: Estimated already saturated. 59 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-05-13) | 80.86 | #325 |
| 2 | GPT-4 (0613) | 77.85 | #491 |
| 3 | GPT-4 Turbo | 77.74 | #404 |
| 4 | GPT-4 Preview (0125) | 76.83 | #486 |
| 5 | Claude 3 Opus (20240229) | 76.55 | #412 |
| 6 | GPT-4 (0314) | 75.67 | #574 |
| 7 | GPT-4 Preview (1106) | 74.96 | #418 |
| 8 | Llama 3 70B Instruct | 67.97 | #634 |
| 9 | Mistral Large | 67.69 | #710 |
| 10 | Qwen 1.5 110B Chat | 66.09 | #691 |
| 11 | Yi 1.5 34B Chat | 64.85 | #800 |
| 12 | Llama 3 8B Instruct | 63.84 | #1129 |
| 13 | Qwen 1.5 72B Chat | 63.47 | #776 |
| 14 | Mixtral 8x22B Instruct (v0.1) | 62.41 | #756 |
| 15 | Mistral Medium | 62.15 | #792 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=magi-hard · How It Works · Data refreshed daily, snapshot 2026-10-20.