Calibration
AA ITBench → HLE w/ tools
HLE w/ tools is estimated from AA ITBench with a linear curve fitted on 5 models measured on both: y = -0.9379·x + 1.0465, R² = 0.82, cross-validated error 4.6 pp. It is used for 6 estimates.
| Estimated model | AA ITBench | HLE w/ tools | Source |
|---|---|---|---|
| Claude Fable 5.1 | 49.5% | 58.2% | estimated ± 4.6 pp, medium confidence |
| Gemini 3.8 Flash | 52.5% | 55.4% | estimated ± 4.6 pp, low confidence |
| GLM-5.2 | 42.7% | 64.6% | estimated ± 4.6 pp, medium confidence |
| GPT-6 Sol | 49.4% | 58.3% | estimated ± 4.6 pp, medium confidence |
| Grok 4.7 | 42.1% | 65.2% | estimated ± 4.6 pp, medium confidence |
| Muse Spark 1.3 | 33.2% | 73.5% | estimated ± 4.6 pp, low confidence |