Calibration
Toolathlon → MCP-Tasks
MCP-Tasks is estimated from Toolathlon with a Michaelis–Menten curve fitted on 5 models measured on both: y = 1.1694·x / (0.26351 + x), R² = 0.43, cross-validated error 7.0 pp. It is used for 9 estimates.
| Estimated model | Toolathlon | MCP-Tasks | Source |
|---|---|---|---|
| GPT-5.6 Sol | 58.0% | 80.4% | estimated ± 7.0 pp, low confidence |
| Gemini 3.5 Flash | 56.5% | 79.7% | estimated ± 7.0 pp, low confidence |
| GPT-5.5 | 55.6% | 79.3% | estimated ± 7.0 pp, low confidence |
| GPT-5.6 Luna | 53.4% | 78.3% | estimated ± 7.0 pp, low confidence |
| GPT-5.6 Terra | 53.1% | 78.2% | estimated ± 7.0 pp, low confidence |
| DeepSeek V4 Pro 0813 | 51.8% | 77.5% | estimated ± 7.0 pp, low confidence |
| Step 3.7 Flash | 49.5% | 76.3% | estimated ± 7.0 pp, low confidence |
| GLM-5.2 | 48.2% | 75.6% | estimated ± 7.0 pp, low confidence |
| DeepSeek V4 Flash 0731 | 47.8% | 75.4% | estimated ± 7.0 pp, low confidence |
Fitted on
| Model | Toolathlon | MCP-Tasks |
|---|---|---|
| Claude Opus 4.5 | 43.5% | 71.8% |
| GLM-5 | 38.0% | 60.8% |
| Kimi K2.5 | 27.8% | 59.1% |
| Qwen3.5 397B | 36.3% | 74.2% |
| Qwen3.6 Plus | 39.8% | 74.1% |