Calibration
Vals LiveCodeBench → AA-SciCode
AA-SciCode is estimated from Vals LiveCodeBench with a offset logistic curve fitted on 37 models measured on both: y = 0.4947 + (0.6532 − 0.4947) / (1 + exp(−83.07·(x − 0.8887))), R² = 0.38, cross-validated error 4.7 pp. It is used for 1 estimate.
| Estimated model | Vals LiveCodeBench | AA-SciCode | Source |
|---|---|---|---|
| GLM-4.5 | 67.4% | 49.5% | estimated ± 4.7 pp, low confidence |
Fitted on
| Model | Vals LiveCodeBench | AA-SciCode |
|---|---|---|
| Claude Fable 5 | 89.8% | 61.0% |
| Claude Fable 5.1 | 90.5% | 63.1% |
| Claude Opus 4.8 | 87.8% | 54.4% |
| Claude Opus 5 | 89.0% | 56.4% |
| Claude Sonnet 5 | 82.4% | 54.3% |
| DeepSeek V4 Flash 0731 | 87.3% | 50.3% |
| DeepSeek V4 Pro 0813 | 87.5% | 51.0% |
| Gemini 3.1 Pro | 88.5% | 58.7% |
| Gemini 3.5 Flash | 87.6% | 53.9% |
| Gemini 3.5 Flash-Lite | 79.0% | 41.3% |
| Gemini 3.6 Flash | 88.1% | 53.4% |
| Gemini 3.7 Flash | 88.7% | 57.2% |
| Gemini 3.8 Flash | 89.5% | 56.6% |
| GLM-5.1 | 81.4% | 44.8% |
| GLM-5.2 | 69.5% | 51.2% |
| GLM-5.3 | 80.5% | 59.0% |
| GPT-5.4 mini | 81.5% | 52.1% |
| GPT-5.4 nano | 84.0% | 47.2% |
| GPT-5.5 | 85.3% | 55.8% |
| GPT-5.6 Sol | 82.6% | 57.1% |
| GPT-5.6 Terra | 85.9% | 55.0% |
| Grok 4.3 | 84.5% | 48.3% |
| Grok 4.5 | 87.4% | 55.0% |
| Grok 4.6 | 88.2% | 56.5% |
| Inkling | 85.5% | 47.0% |
| Inkling-Small | 85.9% | 49.7% |
| Kimi K2.6 | 86.8% | 51.5% |
| Kimi K2.7 Code | 82.1% | 47.8% |
| Kimi K3 | 87.2% | 59.5% |
| Ling 3.0 Flash | 84.0% | 42.0% |
| MiMo-V2.5-Pro | 81.4% | 50.6% |
| MiniMax M2.7 | 79.9% | 50.1% |
| MiniMax M3 | 82.2% | 47.1% |
| Muse Spark 1.1 | 85.9% | 58.8% |
| Nemotron 3 Ultra | 86.0% | 40.3% |
| Qwen3.7 Max | 87.1% | 49.5% |
| Qwen3.8-27B | 84.0% | 46.6% |