Vals AI MMMU: leaderboard
Vals AI run of MMMU Pro, the filtered version of MMMU, using its four option standard split of about 1,700 multiple choice questions that pair images with text across 30 college subjects. Models answer in one pass at temperature zero and the selected letter is matched exactly against the key. Vals says it no longer runs it on new releases, citing saturation; the percentage correct is reported, higher is better.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around January 2027. 93 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 | 90.64 | #9 |
| 2 | Claude Opus 5 | 89.88 | #12 |
| 3 | Claude Fable 5 | 89.31 | #16 |
| 4 | Gemini 3.8 Flash | 89.08 | #21 |
| 5 | Gemini 3.7 Flash | 88.96 | #23 |
| 6 | GPT-5.6 Sol | 88.84 | #19 |
| 7 | Gemini 3.6 Flash | 88.38 | #51 |
| 8 | GPT-5.5 | 88.27 | #31 |
| 9 | Gemini 3.5 Flash | 88.27 | #60 |
| 10 | Gemini 3.1 Pro (Preview) | 88.21 | #58 |
| 11 | Kimi K3 | 88.15 | #30 |
| 12 | Qwen 3.8 Max | 88.03 | #35 |
| 13 | Gemini 3 Flash (Preview) | 87.63 | #82 |
| 14 | Gemini 3 Pro (Preview) | 87.51 | #66 |
| 15 | GPT-5.4 (2026-03-05) | 87.51 | #59 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-mmmu · How It Works · Data refreshed daily, snapshot 2026-10-20.