MMMU Pro
AI benchmark results and model performance
Type: percentage
Recorded results
132
Average score
66.27%
Score range
33% - 86.70%
Leading model
86.70% - GPT-6 Astra
Scores over time
Individual benchmark scores plotted by date.
Model results
| Organisation | Models | Reported | Top score | Info | Self reported | Source |
|---|---|---|---|---|---|---|
| GPT-6 Astra | 03 Sept 2026 | 86.70% | MMMU-Pro (Vision); reasoning effort max; estimated cost per task $0.065; OpenAI-reported in the GPT-6.1 Terra launch post | Yes | Source | |
| GPT-6.1 Sol | 29 Sept 2026 | 86.30% | MMMU-Pro (Vision); reasoning effort max; estimated cost per task $0.013; OpenAI-reported in the GPT-6.1 Terra launch post | Yes | Source | |
| GPT-6.1 Terra | 26 Oct 2026 | 85.50% | MMMU-Pro (Vision); reasoning effort max; estimated cost per task $0.012; OpenAI-reported in the GPT-6.1 Terra launch post | Yes | Source | |
| GPT 5.5 | 23 Apr 2026 | 83.20% | With tools | Yes | Source | |
| GPT 5.4 | 05 Mar 2026 | 82.10% | With tools | Yes | Source | |
| GPT 5.6 Terra | 09 Jul 2026 | 81.70% | MMMU-Pro (Vision); reasoning effort max; estimated cost per task $0.079; OpenAI-reported in the GPT-6.1 Terra launch post | Yes | Source | |
| Kimi K3 | 16 Jul 2026 | 81.60% | Average of three runs; official protocol preserves input order and prepends images to text input. | Yes | Source | |
| Gemini 3 Flash Preview | 17 Dec 2025 | 81.20% | - | Yes | Source | |
| GPT 5 Pro | 07 Aug 2025 | 81.20% | inferred family alias from gpt-5.4 (score=0.4083; benches=19) | Yes | Source | |
| GPT 5 Search API | 14 Oct 2025 | 81.20% | inferred family alias from gpt-5.4 (score=0.3050; benches=19) | Yes | Source | |
| Gemini 3 Pro Image Preview (Nano Banana Pro) | 20 Nov 2025 | 81% | inferred modality/version alias from gemini-3-pro-preview | Yes | Source | |
| Gemini 3 Pro Preview | 18 Nov 2025 | 81% | - | Yes | Source | |
| Gemini 3.1 Pro Preview | 19 Feb 2026 | 80.50% | Search + Python + Browse | Yes | - | |
| Muse Spark | 08 Apr 2026 | 80.40% | Multimodal understanding | Yes | Source | |
| GPT-6 Luna | 22 Sept 2026 | 79.90% | MMMU-Pro (Vision); reasoning effort max; estimated cost per task $0.0021; OpenAI-reported in the GPT-6.1 Terra launch post | Yes | Source | |
| GPT 5.2 Chat | 11 Dec 2025 | 79.50% | inferred alias from gpt-5.2-2025-12-11 | Yes | Source | |
| Qwen 3.5 397B A17B | 16 Feb 2026 | 79% | - | Yes | Source | |
| Qwen 3.6 Plus | 01 Apr 2026 | 78.80% | - | Yes | Source | |
| Kimi K2.5 | 27 Jan 2026 | 78.50% | - | Yes | Source | |
| GPT 5 | 07 Aug 2025 | 78.40% | With Thinking, Pass @ 1 | Yes | Source | |
| Seed 2.0 Pro | 14 Feb 2026 | 78.20% | Seed2 official benchmark table | MMMU-Pro | Yes | Source | |
| GPT 5.4 Mini | 17 Mar 2026 | 78% | With Python | Yes | Source | |
| MiMo V2.5 | 23 Apr 2026 | 77.90% | Image understanding benchmark from official MiMo-V2.5 model card image | Yes | Source | |
| Claude Opus 4.6 | 05 Feb 2026 | 77.30% | With tools | Yes | Source | |
| Gemma 4 31B | 02 Apr 2026 | 76.90% | - | Yes | Source |
