Calibration
Vals GPQA Diamond → HLE w/o tools
HLE w/o tools is estimated from Vals GPQA Diamond with a offset logistic curve fitted on 19 models measured on both: y = 0.3058 + (0.4920 − 0.3058) / (1 + exp(−200.00·(x − 0.8931))), R² = 0.69, cross-validated error 6.6 pp. It is used for 0 estimates.
Fitted on
| Model | Vals GPQA Diamond | HLE w/o tools |
|---|---|---|
| Claude Fable 5.1 | 93.4% | 60.9% |
| Claude Opus 4.8 | 92.4% | 49.8% |
| Claude Opus 5 | 93.4% | 56.3% |
| Claude Sonnet 5 | 88.9% | 43.2% |
| Gemini 3.1 Pro | 95.5% | 45.4% |
| GLM-5.2 | 85.6% | 40.5% |
| GPT-5.4 mini | 83.1% | 28.2% |
| GPT-5.4 nano | 77.5% | 24.3% |
| GPT-5.5 | 93.2% | 41.4% |
| Grok 4.20 | 88.6% | 31.6% |
| Inkling | 87.1% | 30.0% |
| Inkling-Small | 83.6% | 31.6% |
| Kimi K3 | 92.9% | 43.5% |
| MiMo-V2.5-Pro | 82.6% | 34.0% |
| Muse Spark | 89.6% | 42.8% |
| Muse Spark 1.1 | 91.2% | 52.2% |
| Nemotron 3 Ultra | 86.1% | 26.7% |
| Qwen3.8-27B | 88.9% | 30.8% |
| Qwen3.8 Max | 93.7% | 43.6% |