Calibration
Toolathlon → GDPval-AA
GDPval-AA is estimated from Toolathlon with a Michaelis–Menten + offset curve fitted on 19 models measured on both: y = 0.0000 + 2.0000·x / (2.18253 + x), R² = 0.44, cross-validated error 9.6 pp. It is used for 0 estimates.
Fitted on
| Model | Toolathlon | GDPval-AA |
|---|---|---|
| Claude Opus 4.8 | 59.9% | 47.7% |
| DeepSeek V4 Flash 0731 | 47.8% | 46.7% |
| DeepSeek V4 Pro 0813 | 51.8% | 54.5% |
| Gemini 3.5 Flash | 56.5% | 42.2% |
| GLM-5.2 | 48.2% | 43.6% |
| GPT-5.4 | 54.6% | 37.3% |
| GPT-5.4 mini | 42.9% | 25.6% |
| GPT-5.4 nano | 35.5% | 22.4% |
| GPT-5.5 | 55.6% | 42.5% |
| GPT-5.6 Luna | 53.4% | 48.0% |
| GPT-5.6 Sol | 58.0% | 55.4% |
| GPT-5.6 Terra | 53.1% | 47.6% |
| Kimi K2.6 | 50.0% | 26.9% |
| Kimi K2.5 | 27.8% | 17.0% |
| MiniMax M2.7 | 46.3% | 25.5% |
| Muse Spark 1.1 | 75.6% | 35.6% |
| Qwen3.6-35B-A3B | 26.9% | 19.6% |
| Qwen3.6 Plus | 39.8% | 24.5% |
| Step 3.7 Flash | 49.5% | 25.8% |