Calibration
ResearchClawBench → MCP Atlas
MCP Atlas is estimated from ResearchClawBench with a Hill curve fitted on 12 models measured on both: y = 0.0000 + (0.8463 − 0.0000)·x^6.00 / (0.13898^6.00 + x^6.00), R² = 0.54, cross-validated error 13.5 pp. It is used for 3 estimates.
| Estimated model | ResearchClawBench | MCP Atlas | Source |
|---|---|---|---|
| MiMo-V2.5 | 16.9% | 64.6% | estimated ± 13.5 pp, low confidence |
| MiMo-V2-Pro | 15.3% | 54.2% | estimated ± 13.5 pp, low confidence |
| Grok 4.1 | 13.5% | 38.6% | estimated ± 13.5 pp, low confidence |
Fitted on
| Model | ResearchClawBench | MCP Atlas |
|---|---|---|
| Claude Opus 4.8 | 21.1% | 82.2% |
| Gemini 3.5 Flash | 18.0% | 83.6% |
| GLM-5.1 | 18.2% | 71.8% |
| GLM-5.2 | 20.7% | 76.8% |
| GPT-5.4 | 15.3% | 70.6% |
| GPT-5.5 | 17.0% | 75.3% |
| Kimi K2.6 | 18.0% | 55.9% |
| Kimi K2.5 | 14.0% | 29.5% |
| MiniMax M3 | 19.8% | 74.2% |
| Qwen3.5 397B | 14.2% | 46.1% |
| Qwen3.6 Plus | 18.0% | 48.2% |
| Qwen3.7 Max | 18.7% | 76.4% |