Calibration
EBR-bench → ExploitBench
ExploitBench is estimated from EBR-bench with a inverse Michaelis–Menten curve fitted on 5 models measured on both: y = 3.32881·(x − 0.3317) / (0.3317 + 0.6428 − x), R² = 0.98, cross-validated error 6.4 pp. It is used for 17 estimates.
| Estimated model | EBR-bench | ExploitBench | Source |
|---|---|---|---|
| Claude Fable 5.1 | 57.1% | 100.0% | estimated ± 6.4 pp, low confidence |
| Claude Fable 5.2 | 90.5% | 100.0% | estimated ± 6.4 pp, low confidence |
| Claude Haiku 5.5 | 43.8% | 65.9% | estimated ± 6.4 pp, low confidence |
| Claude Fable 5 | 39.5% | 36.3% | estimated ± 6.4 pp, low confidence |
| Claude Opus 4.5 | 14.3% | 0.0% | estimated ± 6.4 pp, low confidence |
| Claude Opus 4.6 | 12.7% | 0.0% | estimated ± 6.4 pp, low confidence |
| Claude Opus 4.7 | 19.0% | 0.0% | estimated ± 6.4 pp, low confidence |
| Claude Sonnet 4.5 | 2.4% | 0.0% | estimated ± 6.4 pp, low confidence |
| Gemini 3.5 Flash | 4.8% | 0.0% | estimated ± 6.4 pp, low confidence |
| GLM-5.2 | 9.5% | 0.0% | estimated ± 6.4 pp, low confidence |
| GPT-5 | 12.7% | 0.0% | estimated ± 6.4 pp, low confidence |
| GPT-5.2 | 23.0% | 0.0% | estimated ± 6.4 pp, low confidence |
| GPT-5.4 | 25.4% | 0.0% | estimated ± 6.4 pp, low confidence |
| Grok 4.6 | 30.5% | 0.0% | estimated ± 6.4 pp, low confidence |
| Kimi K2.6 | 2.4% | 0.0% | estimated ± 6.4 pp, low confidence |
| Muse Wonder | 31.4% | 0.0% | estimated ± 6.4 pp, low confidence |
| Qwen3.7 Max | 9.5% | 0.0% | estimated ± 6.4 pp, low confidence |
Fitted on
| Model | EBR-bench | ExploitBench |
|---|---|---|
| GPT-5.6 Sol | 44.8% | 73.5% |
| GPT-6.1 Sol | 54.3% | 99.7% |
| GPT-6.1 Terra | 78.1% | 97.1% |
| GPT-6.2 Astra | 92.4% | 100.0% |
| GPT-6 Astra | 76.2% | 100.0% |