Calibration
CharXiv → OmniDocBench 1.5
OmniDocBench 1.5 is estimated from CharXiv with a Michaelis–Menten curve fitted on 5 models measured on both: y = 2.0000·x / (1.08154 + x), R² = 0.38, cross-validated error 7.4 pp. It is used for 8 estimates.
| Estimated model | CharXiv | OmniDocBench 1.5 | Source |
|---|---|---|---|
| Claude Mythos 5 | 93.5% | 92.7% | estimated ± 7.4 pp, low confidence |
| Claude Opus 4.8 | 89.9% | 90.8% | estimated ± 7.4 pp, low confidence |
| Gemini 3.1 Flash-Lite | 73.2% | 80.7% | estimated ± 7.4 pp, low confidence |
| GLM-5.3-Flash | 89.4% | 90.5% | estimated ± 7.4 pp, low confidence |
| Muse Spark 1.1 | 88.4% | 89.9% | estimated ± 7.4 pp, low confidence |
| Qwen3.8-Omni-Flash | 91.4% | 91.6% | estimated ± 7.4 pp, low confidence |
| Sakana Fugu | 85.1% | 88.1% | estimated ± 7.4 pp, low confidence |
| Sakana Fugu-Ultra | 86.6% | 88.9% | estimated ± 7.4 pp, low confidence |