Calibration
MCP Atlas → JobBench
JobBench is estimated from MCP Atlas with a inverse Michaelis–Menten curve fitted on 9 models measured on both: y = 0.75626·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.88, cross-validated error 6.8 pp. It is used for 14 estimates.
| Estimated model | MCP Atlas | JobBench | Source |
|---|---|---|---|
| Gemini 3.5 Flash | 83.6% | 54.3% | estimated ± 6.8 pp, medium confidence |
| GLM-5.1 | 71.8% | 42.4% | estimated ± 6.8 pp, medium confidence |
| GLM-5.2 | 76.8% | 47.1% | estimated ± 6.8 pp, medium confidence |
| GPT-5.4 mini | 57.7% | 30.7% | estimated ± 6.8 pp, medium confidence |
| GPT-5.4 nano | 56.1% | 29.5% | estimated ± 6.8 pp, medium confidence |
| Inkling | 74.1% | 44.5% | estimated ± 6.8 pp, medium confidence |
| Kimi K2.7 Code | 76.0% | 46.4% | estimated ± 6.8 pp, medium confidence |
| LLaDA2.2-flash | 46.2% | 22.7% | estimated ± 6.8 pp, medium confidence |
| LongCat-Flash-Lite-Sparse | 45.6% | 22.3% | estimated ± 6.8 pp, medium confidence |
| Muse Glimmer 30B | 75.5% | 45.9% | estimated ± 6.8 pp, medium confidence |
| Qwen3.7 Max | 76.4% | 46.7% | estimated ± 6.8 pp, medium confidence |
| Beam | 78.7% | 49.1% | estimated ± 6.8 pp, medium confidence |
| Solar Open 2 | 58.2% | 31.0% | estimated ± 6.8 pp, medium confidence |
| Solar Pro 4 | 61.4% | 33.5% | estimated ± 6.8 pp, medium confidence |