GPT6 Astra: How General-Purpose LLMs Are Entering the Embodied AI Space

GPT-6 Astra Picks Up a Robot Arm — And Changes the Conversation Around VLA Models OpenAI's latest frontier model wasn't designed to be a robot policy. But early evaluations suggest it might not need to be. --- What Is GPT-6 Astra? GPT-6 Astra was released by OpenAI on September 3, 2026. It is OpenAI's largest training run to date — over 100,000 GPUs at the Stargate site in Texas — and the first release where earlier OpenAI models supervised the training of the newer one. OpenAI called it the world's most intelligent and aligned model. Greg Brockman told reporters it is "not unreasonable to feel that we are now in the AGI era." Beyond the headline benchmarks in reasoning, math, and software engineering, something quieter but arguably more significant happened in the weeks that followed: robotics researchers started handing Astra control of real robot arms. What they found has opened a genuine debate about whether large language models — trained on nothing robot-specific — can serve as competitive robot policies. --- The RoboCurve Experiment: Real Arms, Two Tasks RoboCurve, a public benefit corporation backed by Y Combinator, runs one of the few independent robotics evaluations using real hardware, real tasks, and a human grader scoring every trial. Their September 4, 2026 post is a direct follow-up to their earlier comparison of Claude Fable 5 and Fable 5.1. This time, they gave GPT-6 Astra control of the same pair of I2RT YAM bimanual robot arms under the same open-source…