Zhiyuan’s GO-3 Nano Preview Tops RoboDojo Benchmark, Beats GPT-6 Astra

The headline is the ranking, but the sub-scores are the real story: strength in fine manipulation alongside a weaker open semantic understanding suggests a model optimized for controlled tasks rather than general-purpose reasoning in messy environments. That gap matters for anyone building embodied systems where perception and language grounding are the bottleneck. A single benchmark lead over GPT-6 Astra is a snapshot, not a durable position, and preview-stage results leave open whether the margin holds once the model ships and rivals iterate. Watch whether the semantic-understanding deficit narrows in subsequent evaluations.