Calibration
AA AutomationBench → ExploitGym
ExploitGym is estimated from AA AutomationBench with a Hill curve fitted on 7 models measured on both: y = 0.0561 + (1.2000 − 0.0561)·x^6.00 / (0.87948^6.00 + x^6.00), R² = 0.74, cross-validated error 9.8 pp. It is used for 3 estimates.
| Estimated model | AA AutomationBench | ExploitGym | Source |
|---|---|---|---|
| GPT-6.2 Astra | 77.2% | 41.5% | estimated ± 9.8 pp, low confidence |
| Qwen4 Max | 72.6% | 33.1% | estimated ± 9.8 pp, low confidence |
| GPT-6.2 Sol | 71.3% | 30.9% | estimated ± 9.8 pp, low confidence |
Fitted on
| Model | AA AutomationBench | ExploitGym |
|---|---|---|
| Claude Opus 5.5 | 69.5% | 33.3% |
| Gemini 4 Argon | 77.5% | 39.5% |
| Gemini 4 Carbon | 77.8% | 53.5% |
| GLM-5.3 | 62.2% | 15.0% |
| GPT-6 Luna | 53.2% | 11.6% |
| Kimi K3.1 Pro | 73.3% | 20.7% |
| MiMo-V2.6-Pro | 58.6% | 17.8% |