Calibration
Vals MMLU-Pro → GPQA Diamond
GPQA Diamond is estimated from Vals MMLU-Pro with a linear curve fitted on 23 models measured on both: y = 0.9065·x + 0.1177, R² = 0.59, cross-validated error 2.0 pp. It is used for 2 estimates.
| Estimated model | Vals MMLU-Pro | GPQA Diamond | Source |
|---|---|---|---|
| Muse Spark 1.2 | 88.3% | 91.8% | estimated ± 2.0 pp, high confidence |
| Mistral Medium 3.5 128B | 75.3% | 80.0% | estimated ± 2.0 pp, medium confidence |
Fitted on
| Model | Vals MMLU-Pro | GPQA Diamond |
|---|---|---|
| Claude Opus 4.8 | 89.6% | 93.6% |
| DeepSeek V4 Flash 0731 | 86.2% | 88.1% |
| DeepSeek V4 Pro 0813 | 87.0% | 90.1% |
| Gemini 3.1 Pro | 91.0% | 94.3% |
| Gemini 3.5 Flash | 89.5% | 92.7% |
| GLM-5.1 | 86.9% | 86.2% |
| GLM-5.2 | 86.7% | 91.2% |
| GPT-5.5 | 88.1% | 93.6% |
| GPT-5.6 Luna | 86.0% | 92.3% |
| GPT-5.6 Sol | 89.1% | 94.6% |
| GPT-5.6 Terra | 86.7% | 92.9% |
| Grok 4.20 | 86.3% | 88.5% |
| Inkling | 86.3% | 87.9% |
| Inkling-Small | 85.6% | 89.5% |
| Kimi K2.6 | 87.6% | 90.5% |
| Kimi K3 | 88.0% | 93.5% |
| Ling 3.0 Flash | 82.0% | 85.0% |
| MiniMax M2.7 | 80.4% | 87.0% |
| Muse Spark | 87.3% | 89.5% |
| Nemotron 3 Ultra | 85.8% | 87.0% |
| Qwen3.7 Max | 89.3% | 92.4% |
| Qwen3.8-27B | 84.3% | 89.2% |
| Qwen3.8 Max | 88.6% | 92.6% |