Calibration
GDP.pdf → ExploitGym
ExploitGym is estimated from GDP.pdf with a Hill curve fitted on 7 models measured on both: y = 0.1357 + (1.2000 − 0.1357)·x^6.00 / (0.37912^6.00 + x^6.00), R² = 0.90, cross-validated error 5.0 pp. It is used for 17 estimates.
| Estimated model | GDP.pdf | ExploitGym | Source |
|---|---|---|---|
| Claude Fable 5.1 | 26.2% | 24.0% | estimated ± 5.0 pp, low confidence |
| Claude Haiku 5.5 | 20.8% | 16.4% | estimated ± 5.0 pp, low confidence |
| Claude Opus 5.5 | 26.2% | 24.0% | estimated ± 5.0 pp, low confidence |
| Claude Sonnet 5.5 | 25.8% | 23.2% | estimated ± 5.0 pp, low confidence |
| Gemini 3.8 Flash | 21.0% | 16.6% | estimated ± 5.0 pp, low confidence |
| Gemini 4 Argon | 21.8% | 17.3% | estimated ± 5.0 pp, low confidence |
| GLM-5.3-Flash | 15.4% | 14.0% | estimated ± 5.0 pp, low confidence |
| Grok 4.7 | 20.0% | 15.8% | estimated ± 5.0 pp, low confidence |
| Inkling | 12.8% | 13.7% | estimated ± 5.0 pp, low confidence |
| Kimi K3 | 22.0% | 17.5% | estimated ± 5.0 pp, low confidence |
| MiniMax M3 | 9.8% | 13.6% | estimated ± 5.0 pp, low confidence |
| Mistral Large 4 | 18.6% | 15.0% | estimated ± 5.0 pp, low confidence |
| Muse Glimmer 30B | 10.0% | 13.6% | estimated ± 5.0 pp, low confidence |
| Muse Spark 1.3 | 26.6% | 24.9% | estimated ± 5.0 pp, low confidence |
| Nemotron 3 Ultra | 5.0% | 13.6% | estimated ± 5.0 pp, low confidence |
| Qwen3.8-27B | 16.6% | 14.3% | estimated ± 5.0 pp, low confidence |
| Step 5 Preview | 14.8% | 13.9% | estimated ± 5.0 pp, low confidence |