Calibration
Toolathlon → AA Tau3 Banking
AA Tau3 Banking is estimated from Toolathlon with a offset logistic curve fitted on 6 models measured on both: y = 0.3955 + (1.2000 − 0.3955) / (1 + exp(−200.00·(x − 0.5939))), R² = 0.96, cross-validated error 2.0 pp. It is used for 11 estimates.
| Estimated model | Toolathlon | AA Tau3 Banking | Source |
|---|---|---|---|
| Claude Opus 4.8 | 59.9% | 98.8% | estimated ± 2.0 pp, low confidence |
| Gemini 3.5 Flash | 56.5% | 39.8% | estimated ± 2.0 pp, medium confidence |
| GPT-5.6 Luna | 53.4% | 39.6% | estimated ± 2.0 pp, medium confidence |
| Kimi K2.6 | 50.0% | 39.6% | estimated ± 2.0 pp, medium confidence |
| Step 3.7 Flash | 49.5% | 39.6% | estimated ± 2.0 pp, medium confidence |
| GLM-5.2 | 48.2% | 39.6% | estimated ± 2.0 pp, medium confidence |
| MiniMax M2.7 | 46.3% | 39.6% | estimated ± 2.0 pp, low confidence |
| Claude Opus 4.5 | 43.5% | 39.6% | estimated ± 2.0 pp, low confidence |
| GPT-5.4 mini | 42.9% | 39.6% | estimated ± 2.0 pp, low confidence |
| GLM-5 | 38.0% | 39.6% | estimated ± 2.0 pp, low confidence |
| Qwen3.6 Plus | 39.8% | 39.6% | estimated ± 2.0 pp, low confidence |
Fitted on
| Model | Toolathlon | AA Tau3 Banking |
|---|---|---|
| DeepSeek V4 Flash 0731 | 47.8% | 39.4% |
| DeepSeek V4 Pro 0813 | 51.8% | 39.6% |
| GPT-5.4 | 54.6% | 39.6% |
| GPT-5.5 | 55.6% | 39.0% |
| GPT-5.6 Sol | 58.0% | 44.3% |
| GPT-5.6 Terra | 53.1% | 40.2% |