Calibration
Toolathlon → CyberGym
CyberGym is estimated from Toolathlon with a offset logistic curve fitted on 10 models measured on both: y = 0.4319 + (0.7816 − 0.4319) / (1 + exp(−106.68·(x − 0.4473))), R² = 0.77, cross-validated error 8.6 pp. It is used for 0 estimates.
Fitted on
| Model | Toolathlon | CyberGym |
|---|---|---|
| Claude Opus 4.5 | 43.5% | 50.6% |
| DeepSeek V4 Flash 0731 | 47.8% | 76.7% |
| DeepSeek V4 Pro 0813 | 51.8% | 83.3% |
| GLM-5 | 38.0% | 43.2% |
| GPT-5.4 | 54.6% | 79.0% |
| GPT-5.5 | 55.6% | 81.8% |
| GPT-5.6 Luna | 53.4% | 77.9% |
| GPT-5.6 Sol | 58.0% | 84.5% |
| GPT-5.6 Terra | 53.1% | 81.8% |
| Muse Spark 1.1 | 75.6% | 59.0% |