Calibration
AA AutomationBench → CWE-bench v1
CWE-bench v1 is estimated from AA AutomationBench with a Hill curve fitted on 10 models measured on both: y = 0.3702 + (0.7855 − 0.3702)·x^4.86 / (0.61807^4.86 + x^4.86), R² = 0.75, cross-validated error 6.0 pp. It is used for 15 estimates.
| Estimated model | AA AutomationBench | CWE-bench v1 | Source |
|---|---|---|---|
| Claude Haiku 5.5 | 35.4% | 39.6% | estimated ± 6.0 pp, medium confidence |
| Claude Sonnet 5.5 | 71.8% | 65.0% | estimated ± 6.0 pp, medium confidence |
| Gemini 3.8 Flash | 59.9% | 56.2% | estimated ± 6.0 pp, medium confidence |
| GLM-5.3 | 62.2% | 58.1% | estimated ± 6.0 pp, medium confidence |
| GLM-5.3-Flash | 60.4% | 56.6% | estimated ± 6.0 pp, medium confidence |
| GPT-6.1 Sol | 64.9% | 60.2% | estimated ± 6.0 pp, medium confidence |
| GPT-6 Luna | 53.2% | 50.5% | estimated ± 6.0 pp, medium confidence |
| Kimi K3 | 58.3% | 54.9% | estimated ± 6.0 pp, medium confidence |
| MiMo-V2.6-Pro | 58.6% | 55.1% | estimated ± 6.0 pp, medium confidence |
| MiniMax M3 | 21.3% | 37.2% | estimated ± 6.0 pp, medium confidence |
| Mistral Large 4 | 59.9% | 56.2% | estimated ± 6.0 pp, medium confidence |
| Muse Glimmer 30B | 6.8% | 37.0% | estimated ± 6.0 pp, medium confidence |
| Nemotron 3 Ultra | 3.0% | 37.0% | estimated ± 6.0 pp, low confidence |
| Qwen3.8-27B | 48.2% | 46.6% | estimated ± 6.0 pp, medium confidence |
| Step 5 Preview | 51.0% | 48.7% | estimated ± 6.0 pp, medium confidence |