Calibration
CharXiv → VideoMMMU
VideoMMMU is estimated from CharXiv with a Hill curve fitted on 8 models measured on both: y = 0.8256 + (1.2000 − 0.8256)·x^6.00 / (1.24847^6.00 + x^6.00), R² = 0.57, cross-validated error 1.5 pp. It is used for 16 estimates.
| Estimated model | CharXiv | VideoMMMU | Source |
|---|---|---|---|
| Claude Mythos 5 | 93.5% | 88.2% | estimated ± 1.5 pp, high confidence |
| Claude Opus 4.7 (Adaptive) | 91.0% | 87.4% | estimated ± 1.5 pp, high confidence |
| Claude Opus 4.8 | 89.9% | 87.1% | estimated ± 1.5 pp, high confidence |
| Claude Sonnet 4.6 | 77.4% | 84.6% | estimated ± 1.5 pp, high confidence |
| Claude Sonnet 5 | 88.3% | 86.7% | estimated ± 1.5 pp, high confidence |
| Gemini 3.1 Flash-Lite | 73.2% | 84.0% | estimated ± 1.5 pp, high confidence |
| Gemini 3.7 Flash | 88.7% | 86.8% | estimated ± 1.5 pp, high confidence |
| GLM-5.3-Flash | 89.4% | 87.0% | estimated ± 1.5 pp, high confidence |
| Muse Spark 1.1 | 88.4% | 86.8% | estimated ± 1.5 pp, high confidence |
| Nemotron 3 Nano Omni 30B A3B | 76.3% | 84.4% | estimated ± 1.5 pp, high confidence |
| Qwen3.5-122B-A10B | 77.2% | 84.5% | estimated ± 1.5 pp, high confidence |
| Qwen3.8-27B | 90.2% | 87.2% | estimated ± 1.5 pp, high confidence |
| Qwen3.8-Flash-Next | 90.6% | 87.3% | estimated ± 1.5 pp, high confidence |
| Qwen3.8-Omni-Flash | 91.4% | 87.6% | estimated ± 1.5 pp, high confidence |
| Sakana Fugu | 85.1% | 86.0% | estimated ± 1.5 pp, high confidence |
| Sakana Fugu-Ultra | 86.6% | 86.3% | estimated ± 1.5 pp, high confidence |