OpenAI's GPT-6-Astra, touted as the most intelligent and human-aligned model, has surprised the embodied AI community by acing traditional benchmarks but failing in real-world robot tests.

• OpenAI said GPT-6-Astra is the most intelligent and human-aligned model, but the embodied community found it struggles in real-world robot tests despite perfect benchmark scores.
• The model's performance has sparked debate in the embodied community about the limits of general-purpose large models in physical tasks.