GPT-6 Astra: 92% Navigation, 0% Locomotion—39-Second Inference Latency Is the Physical-AI Wall

The most systematic test yet of a frontier LLM as a direct robot controller finds a hard asymmetry: language and navigation tasks succeed at 82–92% but real-time locomotion fails entirely—not from reasoning gaps but because 39.86-second average inference latency makes physics-based sequential tasks physically impossible.