Moonshot AI's new PerceptionBench isolates raw visual 'seeing' from reasoning, and no frontier model — including the leading GPT-5.6 Sol — breaks 60% accuracy.
Continue to AI University →