New benchmark confirms AI models still perform poorly at visual perception

The PerceptionBench test conducted by Moonshot AI revealed that multimodal AI models fail to achieve 60% accuracy in visual perception. The GPT-5.6 Sol model leads with a narrow margin, highlighting the challenges these models face in understanding images. The findings indicate that many logical errors attributed to reasoning actually occur at the image-reading stage, reflecting a fundamental weakness in how models process visual inputs. This underscores the need for improvements in the deep learning algorithms employed by these models.