benchgap
Calibration

AA Coding Index → PostTrainBench v1.1

PostTrainBench v1.1 is estimated from AA Coding Index with a Hill curve fitted on 12 models measured on both: y = 0.0813 + (1.2000 − 0.0813)·x^6.00 / (0.93681^6.00 + x^6.00), R² = 0.56, cross-validated error 5.7 pp. It is used for 6 estimates.

Estimated modelAA Coding IndexPostTrainBench v1.1Source
Qwen3.8 Max Preview71.8%27.0%estimated ± 5.7 pp, medium confidence
Quasar 438B61.2%16.2%estimated ± 5.7 pp, low confidence
Apodex 1.160.8%15.9%estimated ± 5.7 pp, low confidence
Apodex 1.1 Mini60.8%15.9%estimated ± 5.7 pp, low confidence
Hy358.8%14.6%estimated ± 5.7 pp, low confidence
Hy3 Preview58.8%14.6%estimated ± 5.7 pp, low confidence

Fitted on

ModelAA Coding IndexPostTrainBench v1.1
Claude Fable 5.181.6%40.2%
Claude Opus 4.773.6%28.6%
Claude Opus 4.874.3%32.9%
Claude Opus 578.0%35.0%
Gemini 3.1 Pro68.8%22.0%
GLM-5.268.8%31.7%
GPT-5.471.1%19.0%
GPT-5.574.9%27.2%
GPT-5.6 Sol77.4%36.2%
GPT-6 Astra76.9%44.3%
Grok 4.572.5%23.5%
Kimi K376.2%32.0%