Calibration
τ³-bench results → MCP-Tasks
MCP-Tasks is estimated from τ³-bench results with a Hill curve fitted on 5 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.65083^6.00 + x^6.00), R² = 0.83, cross-validated error 3.6 pp. It is used for 6 estimates.
| Estimated model | τ³-bench results | MCP-Tasks | Source |
|---|---|---|---|
| MiMo-V2.5-Pro | 72.9% | 79.7% | estimated ± 3.6 pp, low confidence |
| Nemotron 3 Ultra | 70.9% | 75.1% | estimated ± 3.6 pp, low confidence |
| GLM-5.1 | 70.6% | 74.4% | estimated ± 3.6 pp, medium confidence |
| MiniMax M2.7 | 67.6% | 66.8% | estimated ± 3.6 pp, medium confidence |
| Pokee-Isaac 28B | 66.2% | 63.1% | estimated ± 3.6 pp, medium confidence |
| Granite 4.2 30B | 62.0% | 51.3% | estimated ± 3.6 pp, low confidence |
Fitted on
| Model | τ³-bench results | MCP-Tasks |
|---|---|---|
| Claude Opus 4.5 | 70.2% | 71.8% |
| GLM-5 | 65.6% | 60.8% |
| Kimi K2.5 | 65.7% | 59.1% |
| Qwen3.5 397B | 68.4% | 74.2% |
| Qwen3.6 Plus | 70.7% | 74.1% |