GPT-6.2 Astra tops Blueprint-Bench 2 physical 3D benchmark
Surpassing a human baseline on floorplan reconstruction suggests multimodal models are moving from object recognition toward genuine spatial inference, which matters for robotics and embodied applications. Teams building indoor navigation, real-estate tech, or simulation tools should track which frontier models can ingest raw interior photos and produce structured spatial outputs. Developers can benchmark their own spatial pipelines against Blueprint-Bench 2 and compare Astra's ranking with Sol and other frontier models on the same leaderboard.