Calibration
AA-MMMU-Pro → MedXpertQA (MM)
MedXpertQA (MM) is estimated from AA-MMMU-Pro with a offset logistic curve fitted on 6 models measured on both: y = 0.0000 + (0.7792 − 0.0000) / (1 + exp(−37.57·(x − 0.6832))), R² = 0.93, cross-validated error 6.5 pp. It is used for 64 estimates.
| Estimated model | AA-MMMU-Pro | MedXpertQA (MM) | Source |
|---|---|---|---|
| Apodex 1.1 | 79.2% | 76.6% | estimated ± 6.5 pp, low confidence |
| Apodex 1.1 Mini | 79.2% | 76.6% | estimated ± 6.5 pp, low confidence |
| Claude 3 Haiku | 30.8% | 0.0% | estimated ± 6.5 pp, low confidence |
| Claude 4.1 Opus Thinking | 67.9% | 35.9% | estimated ± 6.5 pp, low confidence |
| Claude 4 Sonnet | 62.4% | 7.6% | estimated ± 6.5 pp, low confidence |
| Claude Opus 4.5 Thinking | 74.0% | 69.7% | estimated ± 6.5 pp, low confidence |
| Claude Opus 4.6 (Adaptive) | 75.4% | 72.8% | estimated ± 6.5 pp, low confidence |
| Claude Opus 4.7 | 76.4% | 74.3% | estimated ± 6.5 pp, low confidence |
| Claude Opus 5 | 84.7% | 77.8% | estimated ± 6.5 pp, low confidence |
| Claude Opus 5.5 | 87.7% | 77.9% | estimated ± 6.5 pp, low confidence |
| DeepSeek V4.1 Flash | 77.0% | 75.0% | estimated ± 6.5 pp, low confidence |
| Gemini 1.5 Pro | 55.0% | 0.5% | estimated ± 6.5 pp, low confidence |
| Gemini 2.5 Flash | 65.5% | 20.1% | estimated ± 6.5 pp, low confidence |
| Gemini 2.5 Pro | 74.9% | 71.9% | estimated ± 6.5 pp, low confidence |
| Gemini 3.5 Flash-Lite | 79.0% | 76.5% | estimated ± 6.5 pp, low confidence |
| Gemini 3.6 Flash | 83.2% | 77.6% | estimated ± 6.5 pp, low confidence |
| Gemini 3.8 Flash | 85.6% | 77.8% | estimated ± 6.5 pp, low confidence |
| Gemini 3 Flash | 78.6% | 76.3% | estimated ± 6.5 pp, low confidence |
| Gemma 3 27B | 48.0% | 0.0% | estimated ± 6.5 pp, low confidence |
| Gemma 4 E2B | 44.6% | 0.0% | estimated ± 6.5 pp, low confidence |
| Gemma 4 E4B | 51.4% | 0.1% | estimated ± 6.5 pp, low confidence |
| GLM-5V-Turbo | 72.8% | 65.7% | estimated ± 6.5 pp, low confidence |
| GPT-4.1 | 61.2% | 5.0% | estimated ± 6.5 pp, low confidence |
| GPT-4.1 mini | 58.7% | 2.0% | estimated ± 6.5 pp, low confidence |
| GPT-4.1 nano | 40.1% | 0.0% | estimated ± 6.5 pp, low confidence |
| GPT-4o mini | 41.5% | 0.0% | estimated ± 6.5 pp, low confidence |
| GPT-5.1 | 75.5% | 73.0% | estimated ± 6.5 pp, low confidence |
| GPT-5.1-Codex | 72.5% | 64.5% | estimated ± 6.5 pp, low confidence |
| GPT-5.1-Codex-Max | 72.5% | 64.5% | estimated ± 6.5 pp, low confidence |
| GPT-5.2-Codex | 76.3% | 74.2% | estimated ± 6.5 pp, low confidence |
| GPT-5.3 Codex | 78.5% | 76.3% | estimated ± 6.5 pp, low confidence |
| GPT-5 (high) | 74.2% | 70.2% | estimated ± 6.5 pp, low confidence |
| GPT-5 (medium) | 74.3% | 70.5% | estimated ± 6.5 pp, low confidence |
| GPT-6.1 Sol | 86.0% | 77.8% | estimated ± 6.5 pp, low confidence |
| GPT-6 Astra | 86.9% | 77.8% | estimated ± 6.5 pp, low confidence |
| GPT-6 Luna | 79.7% | 76.9% | estimated ± 6.5 pp, low confidence |
| GPT-6 Sol | 82.9% | 77.6% | estimated ± 6.5 pp, low confidence |
| Grok 4 | 68.8% | 42.5% | estimated ± 6.5 pp, low confidence |
| Grok 4.1 Fast | 48.4% | 0.0% | estimated ± 6.5 pp, low confidence |
| Grok 4.1 Fast (Reasoning) | 63.3% | 10.3% | estimated ± 6.5 pp, low confidence |
| Grok 4.5 | 80.4% | 77.1% | estimated ± 6.5 pp, low confidence |
| Grok 4 Fast (Reasoning) | 61.8% | 6.2% | estimated ± 6.5 pp, low confidence |
| LFM2.5-VL-1.6B-Extract | 26.5% | 0.0% | estimated ± 6.5 pp, low confidence |
| Ling 3.0 Flash VL | 79.0% | 76.5% | estimated ± 6.5 pp, low confidence |
| Llama 4 Maverick | 62.1% | 6.9% | estimated ± 6.5 pp, low confidence |
| Llama 4 Scout | 52.9% | 0.2% | estimated ± 6.5 pp, low confidence |
| MiMo-V2.6-Flash | 73.1% | 66.8% | estimated ± 6.5 pp, low confidence |
| MiMo-V2-Omni | 69.9% | 50.2% | estimated ± 6.5 pp, low confidence |
| Mistral Large 3 | 55.7% | 0.7% | estimated ± 6.5 pp, low confidence |
| Mistral Large 4 | 76.4% | 74.3% | estimated ± 6.5 pp, low confidence |
| Mistral Medium 3 | 53.0% | 0.2% | estimated ± 6.5 pp, low confidence |
| Mistral Medium 3.5 128B | 64.9% | 16.9% | estimated ± 6.5 pp, low confidence |
| Mistral Small 4 | 56.8% | 1.0% | estimated ± 6.5 pp, low confidence |
| Mistral Small 4 (Reasoning) | 56.8% | 1.0% | estimated ± 6.5 pp, low confidence |
| Nova Pro | 44.3% | 0.0% | estimated ± 6.5 pp, low confidence |
| o3 | 70.1% | 51.5% | estimated ± 6.5 pp, low confidence |
| Phi-4 Multimodal Instruct | 14.5% | 0.0% | estimated ± 6.5 pp, low confidence |
| Qwen3.5-27B | 75.0% | 72.1% | estimated ± 6.5 pp, low confidence |
| Qwen3.5-35B-A3B | 72.7% | 65.3% | estimated ± 6.5 pp, low confidence |
| Qwen3.5 397B (Reasoning) | 52.7% | 0.2% | estimated ± 6.5 pp, low confidence |
| Qwen3.8 Max Preview | 82.8% | 77.6% | estimated ± 6.5 pp, low confidence |
| Qwen3-Omni-30B-A3B-Instruct | 55.5% | 0.6% | estimated ± 6.5 pp, low confidence |
| Qwen3-Omni-30B-A3B-Thinking | 60.2% | 3.5% | estimated ± 6.5 pp, low confidence |
| Step 3.7 Flash | 75.3% | 72.6% | estimated ± 6.5 pp, low confidence |