Calibration
AA Coding Index → SciCode
SciCode is estimated from AA Coding Index with a linear curve fitted on 16 models measured on both: y = 0.4345·x + 0.2311, R² = 0.74, cross-validated error 4.7 pp. It is used for 14 estimates.
| Estimated model | AA Coding Index | SciCode | Source |
|---|---|---|---|
| Claude 3 Opus | 19.5% | 31.6% | estimated ± 4.7 pp, high confidence |
| Gemini 1.5 Pro | 23.6% | 33.4% | estimated ± 4.7 pp, high confidence |
| Gemma 4 12B | 31.0% | 36.6% | estimated ± 4.7 pp, high confidence |
| Gemma 4 E2B | 7.2% | 26.3% | estimated ± 4.7 pp, medium confidence |
| GPT-4.1 nano | 11.1% | 28.0% | estimated ± 4.7 pp, medium confidence |
| GPT-4 Turbo | 21.5% | 32.4% | estimated ± 4.7 pp, high confidence |
| GPT-4o mini | 11.4% | 28.1% | estimated ± 4.7 pp, medium confidence |
| GPT-5.1 | 49.4% | 44.6% | estimated ± 4.7 pp, high confidence |
| GPT-5.4 | 71.1% | 54.0% | estimated ± 4.7 pp, medium confidence |
| GPT-5 (high) | 37.8% | 39.5% | estimated ± 4.7 pp, high confidence |
| K-Exaone | 32.1% | 37.1% | estimated ± 4.7 pp, high confidence |
| o1 | 39.7% | 40.4% | estimated ± 4.7 pp, high confidence |
| o1-preview | 34.1% | 37.9% | estimated ± 4.7 pp, high confidence |
| Ultravox v0.6 Llama 3.3 70B | 11.9% | 28.3% | estimated ± 4.7 pp, medium confidence |