Calibration
AA-MMMU-Pro → CharXiv w/o tools
CharXiv w/o tools is estimated from AA-MMMU-Pro with a Michaelis–Menten curve fitted on 9 models measured on both: y = 2.0000·x / (1.13496 + x), R² = 0.62, cross-validated error 2.4 pp. It is used for 61 estimates.
| Estimated model | AA-MMMU-Pro | CharXiv w/o tools | Source |
|---|---|---|---|
| Apodex 1.1 | 79.2% | 82.2% | estimated ± 2.4 pp, high confidence |
| Apodex 1.1 Mini | 79.2% | 82.2% | estimated ± 2.4 pp, high confidence |
| Claude 3 Haiku | 30.8% | 42.7% | estimated ± 2.4 pp, medium confidence |
| Claude 4.1 Opus Thinking | 67.9% | 74.9% | estimated ± 2.4 pp, medium confidence |
| Claude 4 Sonnet | 62.4% | 71.0% | estimated ± 2.4 pp, medium confidence |
| Claude Opus 4.5 Thinking | 74.0% | 78.9% | estimated ± 2.4 pp, high confidence |
| Claude Opus 4.6 (Adaptive) | 75.4% | 79.8% | estimated ± 2.4 pp, high confidence |
| Claude Opus 4.7 | 76.4% | 80.5% | estimated ± 2.4 pp, high confidence |
| Claude Opus 5 | 84.7% | 85.5% | estimated ± 2.4 pp, high confidence |
| Claude Opus 5.5 | 87.7% | 87.2% | estimated ± 2.4 pp, medium confidence |
| DeepSeek V4.1 Flash | 77.0% | 80.8% | estimated ± 2.4 pp, high confidence |
| Gemini 1.5 Pro | 55.0% | 65.3% | estimated ± 2.4 pp, medium confidence |
| Gemini 2.5 Flash | 65.5% | 73.2% | estimated ± 2.4 pp, medium confidence |
| Gemini 2.5 Pro | 74.9% | 79.5% | estimated ± 2.4 pp, high confidence |
| Gemini 3.5 Flash-Lite | 79.0% | 82.1% | estimated ± 2.4 pp, high confidence |
| Gemini 3.6 Flash | 83.2% | 84.6% | estimated ± 2.4 pp, high confidence |
| Gemini 3 Flash | 78.6% | 81.8% | estimated ± 2.4 pp, high confidence |
| Gemma 3 27B | 48.0% | 59.4% | estimated ± 2.4 pp, medium confidence |
| Gemma 4 E2B | 44.6% | 56.4% | estimated ± 2.4 pp, medium confidence |
| Gemma 4 E4B | 51.4% | 62.3% | estimated ± 2.4 pp, medium confidence |
| GLM-5V-Turbo | 72.8% | 78.2% | estimated ± 2.4 pp, medium confidence |
| GPT-4.1 | 61.2% | 70.1% | estimated ± 2.4 pp, medium confidence |
| GPT-4.1 mini | 58.7% | 68.2% | estimated ± 2.4 pp, medium confidence |
| GPT-4.1 nano | 40.1% | 52.2% | estimated ± 2.4 pp, medium confidence |
| GPT-4o mini | 41.5% | 53.5% | estimated ± 2.4 pp, medium confidence |
| GPT-5.1 | 75.5% | 79.9% | estimated ± 2.4 pp, high confidence |
| GPT-5.1-Codex | 72.5% | 78.0% | estimated ± 2.4 pp, medium confidence |
| GPT-5.1-Codex-Max | 72.5% | 78.0% | estimated ± 2.4 pp, medium confidence |
| GPT-5.2-Codex | 76.3% | 80.4% | estimated ± 2.4 pp, high confidence |
| GPT-5.3 Codex | 78.5% | 81.8% | estimated ± 2.4 pp, high confidence |
| GPT-5 (high) | 74.2% | 79.1% | estimated ± 2.4 pp, high confidence |
| GPT-5 (medium) | 74.3% | 79.1% | estimated ± 2.4 pp, high confidence |
| GPT-6.1 Sol | 86.0% | 86.2% | estimated ± 2.4 pp, medium confidence |
| GPT-6 Astra | 86.9% | 86.7% | estimated ± 2.4 pp, medium confidence |
| GPT-6 Luna | 79.7% | 82.5% | estimated ± 2.4 pp, high confidence |
| GPT-6 Sol | 82.9% | 84.4% | estimated ± 2.4 pp, high confidence |
| Grok 4 | 68.8% | 75.5% | estimated ± 2.4 pp, medium confidence |
| Grok 4.1 Fast | 48.4% | 59.8% | estimated ± 2.4 pp, medium confidence |
| Grok 4.1 Fast (Reasoning) | 63.3% | 71.6% | estimated ± 2.4 pp, medium confidence |
| Grok 4.5 | 80.4% | 82.9% | estimated ± 2.4 pp, high confidence |
| Grok 4 Fast (Reasoning) | 61.8% | 70.5% | estimated ± 2.4 pp, medium confidence |
| LFM2.5-VL-1.6B-Extract | 26.5% | 37.9% | estimated ± 2.4 pp, medium confidence |
| Ling 3.0 Flash VL | 79.0% | 82.1% | estimated ± 2.4 pp, high confidence |
| Llama 4 Maverick | 62.1% | 70.7% | estimated ± 2.4 pp, medium confidence |
| Llama 4 Scout | 52.9% | 63.6% | estimated ± 2.4 pp, medium confidence |
| MiMo-V2.6-Flash | 73.1% | 78.4% | estimated ± 2.4 pp, medium confidence |
| MiMo-V2-Omni | 69.9% | 76.2% | estimated ± 2.4 pp, medium confidence |
| Mistral Large 3 | 55.7% | 65.8% | estimated ± 2.4 pp, medium confidence |
| Mistral Large 4 | 76.4% | 80.5% | estimated ± 2.4 pp, high confidence |
| Mistral Medium 3 | 53.0% | 63.7% | estimated ± 2.4 pp, medium confidence |
| Mistral Medium 3.5 128B | 64.9% | 72.8% | estimated ± 2.4 pp, medium confidence |
| Mistral Small 4 | 56.8% | 66.7% | estimated ± 2.4 pp, medium confidence |
| Mistral Small 4 (Reasoning) | 56.8% | 66.7% | estimated ± 2.4 pp, medium confidence |
| Nova Pro | 44.3% | 56.1% | estimated ± 2.4 pp, medium confidence |
| o3 | 70.1% | 76.4% | estimated ± 2.4 pp, medium confidence |
| Phi-4 Multimodal Instruct | 14.5% | 22.7% | estimated ± 2.4 pp, medium confidence |
| Qwen3.5 397B (Reasoning) | 52.7% | 63.4% | estimated ± 2.4 pp, medium confidence |
| Qwen3.8 Max Preview | 82.8% | 84.4% | estimated ± 2.4 pp, high confidence |
| Qwen3-Omni-30B-A3B-Instruct | 55.5% | 65.7% | estimated ± 2.4 pp, medium confidence |
| Qwen3-Omni-30B-A3B-Thinking | 60.2% | 69.3% | estimated ± 2.4 pp, medium confidence |
| Step 3.7 Flash | 75.3% | 79.8% | estimated ± 2.4 pp, high confidence |