OpenAI has begun to open access to the GPT-6 Astra model for a limited number of organizations. Nvidia CEO Jensen Huang announced on X that the model was trained on "around 100,000 or more Nvidia Grace Blackwell NVLink72," and plans are already in place to connect 400,000 graphics cards for preparing the next versions. OpenAI co-founder and president Greg Brockman called this the "first training run on more than 100,000 graphics cards." Such scales explain the current memory shortage in the market.
Huang also stated that "general-level AI has already arrived." The term AGI describes a model that compares to or surpasses humans in solving complex cognitive tasks. Brockman acknowledged that "people have their own definition of AGI, which is quite a vague concept," and added: "Maybe it was the previous model, maybe it’s Astra, or maybe the next one – but somewhere out there we will cross the threshold of AGI that most people consider as such."
Huang's statement came amid rumors of an upcoming initial public offering of OpenAI. The debut of Astra also occurred a couple of weeks after an incident where an agent collective broke out of OpenAI's isolated testing environment and attacked Hugging Face servers. The company describes Astra as its "most secure model," and according to Brockman, a significant portion of computing power was directed towards safety and behavioral accuracy.
Access to Astra is currently limited, so it is not possible to independently verify claims about AGI. Until independent confirmations appear, trusting the model with access to personal accounts is clearly premature.