benchgap
Calibration

AA Coding Index → PostTrainBench

PostTrainBench is estimated from AA Coding Index with a Hill curve fitted on 12 models measured on both: y = 0.0267 + (1.2000 − 0.0267)·x^6.00 / (0.91418^6.00 + x^6.00), R² = 0.47, cross-validated error 5.9 pp. It is used for 4 estimates.

Estimated modelAA Coding IndexPostTrainBenchSource
Qwen3.8 Max Preview71.8%25.0%estimated ± 5.9 pp, low confidence
Quasar 438B61.2%12.4%estimated ± 5.9 pp, low confidence
Apodex 1.160.8%12.0%estimated ± 5.9 pp, low confidence
Apodex 1.1 Mini60.8%12.0%estimated ± 5.9 pp, low confidence

Fitted on

ModelAA Coding IndexPostTrainBench
Claude Opus 4.773.6%24.3%
Claude Opus 4.874.3%31.0%
Claude Opus 578.0%35.7%
Gemini 3.1 Pro68.8%18.1%
GLM-5.268.8%30.4%
GLM-5.374.8%34.5%
GPT-5.471.1%17.8%
GPT-5.574.9%23.9%
GPT-5.6 Sol77.4%31.0%
GPT-6 Astra76.9%41.9%
Grok 4.572.5%22.6%
Kimi K376.2%31.5%