AI MODELS FAIL AT BASIC VISUAL PERCEPTION

PerceptionBench, Moonshot AI's new evaluation tool, isolates visual perception capabilities from logical reasoning to test how well multimodal AI models can actually interpret images. The benchmark exposes a critical weakness: even the leading performer, GPT-5.6 Sol, barely outpaces competitors. The results challenge assumptions about AI reasoning abilities. Developers often attribute model errors to flawed logic, but PerceptionBench demonstrates that fundamental visual comprehension fails before reasoning even begins. This distinction is crucial for identifying where improvements are needed. The benchmark addresses a gap in AI evaluation. While existing tests measure overall multimodal performance, they conflate visual perception with reasoning skills, masking specific deficiencies. PerceptionBench separates these functions to provide clearer insights. These findings highlight remaining limitations in AI vision systems despite recent advances in multimodal models. The narrow performance gap between leading models suggests the field faces a collective challenge in advancing visual perception capabilities.