New BabyVision Benchmark Shows Top AI Trails Toddlers on Visual Tasks

Frontier multimodal AI models score below the average three-year-old on basic, language-free visual reasoning tasks, according to a new benchmark called BabyVision. On the 388-task evaluation, the top performer, Gemini3-Pro-Preview, reached only 49.7% accuracy, far short of the 94.1% posted by adults and roughly 20 points below six-year-olds.