benchgap
Calibration

BabaIsBench → Mystery Game Puzzles

Mystery Game Puzzles is estimated from BabaIsBench with a offset logistic curve fitted on 26 models measured on both: y = 0.2594 + (1.2000 − 0.2594) / (1 + exp(−65.50·(x − 1.0113))), R² = 0.61, cross-validated error 12.4 pp. It is used for 1 estimate.

Estimated modelBabaIsBenchMystery Game PuzzlesSource
Qwen3.8-27B21.0%25.9%estimated ± 12.4 pp, low confidence

Fitted on

ModelBabaIsBenchMystery Game Puzzles
Claude Fable 5100.0%52.0%
Claude Fable 5.1100.0%58.0%
Claude Opus 4.896.0%36.0%
Claude Opus 5100.0%59.0%
Claude Sonnet 567.0%35.0%
DeepSeek V4 Flash 073175.0%34.0%
Gemini 3.1 Pro92.0%34.0%
Gemini 3.5 Flash88.0%32.0%
Gemini 3.6 Flash88.0%30.0%
Gemini 3.7 Flash100.0%37.0%
Gemini 3.8 Flash100.0%47.0%
GLM-5.296.0%15.0%
GLM-5.396.0%33.0%
GLM-5.3-Flash67.0%8.0%
GPT-5.5100.0%56.0%
GPT-5.6 Luna63.0%21.0%
GPT-5.6 Sol100.0%58.0%
GPT-5.6 Terra88.0%35.0%
GPT-6 Astra100.0%84.0%
Grok 4.696.0%34.0%
Kimi K396.0%26.0%
MiniMax M346.0%8.0%
Muse Spark 1.346.0%25.0%
Qwen3.6-27B29.0%7.0%
Qwen3.7 Max25.0%32.0%
Qwen3.8 Max58.0%38.0%