benchgap
Calibration

Vals GPQA Diamond → ARC-AGI-1

ARC-AGI-1 is estimated from Vals GPQA Diamond with a Hill curve fitted on 21 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.76014^6.00 + x^6.00), R² = 0.76, cross-validated error 5.9 pp. It is used for 14 estimates.

Estimated modelVals GPQA DiamondARC-AGI-1Source
MiniMax M392.7%92.0%estimated ± 5.9 pp, medium confidence
Grok 4.391.4%90.2%estimated ± 5.9 pp, medium confidence
GLM-5.388.1%85.0%estimated ± 5.9 pp, medium confidence
Gemini 3 Flash87.9%84.6%estimated ± 5.9 pp, medium confidence
Qwen3.6 Plus87.4%83.8%estimated ± 5.9 pp, medium confidence
GLM-5.3-Flash86.4%82.0%estimated ± 5.9 pp, medium confidence
Gemini 3.5 Flash-Lite83.8%77.1%estimated ± 5.9 pp, medium confidence
Qwen3.5 Flash82.8%75.1%estimated ± 5.9 pp, medium confidence
MiMo-V2.581.6%72.6%estimated ± 5.9 pp, medium confidence
Gemini 3.1 Flash-Lite81.1%71.5%estimated ± 5.9 pp, medium confidence
GLM-4.780.1%69.3%estimated ± 5.9 pp, medium confidence
GLM-4.674.5%56.4%estimated ± 5.9 pp, low confidence
Claude Haiku 4.572.2%50.8%estimated ± 5.9 pp, low confidence
GLM-4.572.2%50.8%estimated ± 5.9 pp, low confidence

Fitted on

ModelVals GPQA DiamondARC-AGI-1
Claude Fable 593.2%98.5%
Claude Fable 5.193.4%97.5%
Claude Opus 4.892.4%92.5%
Claude Opus 593.4%97.5%
Claude Sonnet 4.685.6%86.0%
DeepSeek V4 Flash 073189.9%89.0%
DeepSeek V4 Pro 081392.4%90.0%
Gemini 3.5 Flash92.7%92.5%
Gemini 3.6 Flash93.4%91.2%
Gemini 3.7 Flash93.9%95.5%
Gemini 3.8 Flash94.4%98.5%
GPT-5.4 mini83.1%63.7%
GPT-5.4 nano77.5%51.5%
GPT-5.593.2%95.0%
GPT-5.6 Luna91.7%88.0%
GPT-5.6 Sol95.2%96.5%
GPT-5.6 Terra90.9%96.5%
Grok 4.592.9%85.7%
Grok 4.694.7%87.0%
Inkling-Small83.6%84.0%
Kimi K392.9%94.5%