Elorian and Scale AI released Humanity’s Sixth Sense, a benchmark of 522 visual reasoning tasks. It tests temporal, physical, social and contextual inference from images and videos. Humans score 93.1% on average. Frontier models score below 55%, with GPT 6 Astra leading at 53.6%. 94% of model failures stem from perception or inference, not reasoning. Social understanding is the weakest domain for most models. Video tasks are harder than still images for nearly all models. Increasing reasoning effort sometimes lowers scores. Agent tools modestly improve performance by enabling second looks. The paper and dataset are publicly available.

• 522 tasks in HSS
• 93.1% human baseline pass@1
• 53.6% GPT 6 Astra pass@1
• 25 models from 8 companies