benchgap
Calibration

Mystery Game Puzzles → MaxxBench

MaxxBench is estimated from Mystery Game Puzzles with a Hill curve fitted on 13 models measured on both: y = 0.4718 + (0.7734 − 0.4718)·x^6.00 / (0.54996^6.00 + x^6.00), R² = 0.65, cross-validated error 8.8 pp. It is used for 20 estimates.

Estimated modelMystery Game PuzzlesMaxxBenchSource
o329.0%47.8%estimated ± 8.8 pp, medium confidence
Qwen3.5 Flash20.0%47.3%estimated ± 8.8 pp, medium confidence
Qwen 3.6 Max (preview)19.0%47.2%estimated ± 8.8 pp, medium confidence
Qwen3.5 397B18.0%47.2%estimated ± 8.8 pp, medium confidence
Claude Sonnet 4.517.0%47.2%estimated ± 8.8 pp, medium confidence
Qwen3.5-122B-A10B17.0%47.2%estimated ± 8.8 pp, medium confidence
Qwen3.5 Plus17.0%47.2%estimated ± 8.8 pp, medium confidence
Qwen3.7 Plus17.0%47.2%estimated ± 8.8 pp, medium confidence
Claude Sonnet 4.616.0%47.2%estimated ± 8.8 pp, medium confidence
GPT-5.116.0%47.2%estimated ± 8.8 pp, medium confidence
Qwen3.7 Flash15.0%47.2%estimated ± 8.8 pp, medium confidence
GPT-4o mini12.0%47.2%estimated ± 8.8 pp, medium confidence
Qwen3.6 Plus12.0%47.2%estimated ± 8.8 pp, medium confidence
GPT-4.1 mini7.0%47.2%estimated ± 8.8 pp, medium confidence
GPT-5.4 mini7.0%47.2%estimated ± 8.8 pp, medium confidence
o3-mini7.0%47.2%estimated ± 8.8 pp, medium confidence
GPT-5.4 nano5.0%47.2%estimated ± 8.8 pp, low confidence
o4-mini5.0%47.2%estimated ± 8.8 pp, low confidence
Qwen3 Max5.0%47.2%estimated ± 8.8 pp, low confidence
GPT-OSS 120B0.0%47.2%estimated ± 8.8 pp, low confidence

Fitted on

ModelMystery Game PuzzlesMaxxBench
Claude Fable 552.0%63.5%
Claude Opus 559.0%63.5%
Claude Opus 5.571.0%76.4%
Claude Sonnet 5.565.0%66.9%
DeepSeek V4.1 Flash 101749.0%55.6%
DeepSeek V4 Flash 073134.0%28.1%
Gemini 3.8 Flash47.0%53.4%
GLM-5.333.0%56.7%
GLM-5.3-Flash8.0%55.1%
GPT-5.6 Luna21.0%46.1%
GPT-5.6 Sol58.0%69.1%
GPT-6 Astra84.0%71.9%
GPT-6 Luna7.0%51.7%