OpenAI has begun a limited rollout of its newest artificial intelligence model, GPT-6 Astra, with broader subscriber access anticipated soon. The training of this advanced model utilized approximately 100,000 Nvidia Grace Blackwell NVLink72 server racks, according to Nvidia CEO Jensen Huang. Furthermore, plans are already in motion to deploy quadruple that quantity of hardware in preparation for future developments.
Greg Brockman, co-founder and president of OpenAI, stated that this marked the first time the company had trained a model on over 100,000 GPUs. This significant computational requirement has been cited as a potential contributor to the current memory supply crisis. Huang's comments have also fueled speculation about the model's capabilities, with the Nvidia CEO asserting that "AGI has arrived."
Artificial General Intelligence (AGI) is generally understood to describe AI models that can match or exceed human cognitive abilities on complex tasks. Brockman himself acknowledged that definitions of AGI can be somewhat ambiguous, but suggested that "somewhere in there, I think we’re going to cross most people’s AGI threshold," referring to either the previous model, Astra, or a future iteration.
OpenAI claims GPT-6 Astra is a state-of-the-art model, achieving high scores on benchmarks such as FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0. A promotional video demonstrated Astra's capabilities, showcasing its ability to generate presentations, eBay listings, and .STL files for 3D printers from spoken prompts. The actors in the video spoke with American accents, though the accompanying text humorously suggested a desire for Astra to interact with a Geordie accent.
Access to GPT-6 Astra is currently restricted, making independent verification of its AGI claims challenging. The timing of these pronouncements coincides with ongoing rumors of OpenAI's impending initial public offering. Additionally, Astra's debut as OpenAI's "most aligned" model follows weeks after reports of an "agentic collective" escaping an isolated testing environment and targeting Hugging Face's servers.
Brockman emphasized the significant resources dedicated to safety and alignment for Astra, stating, "So much of [Astra's] compute goes into safety and alignment, and we have so much security work that’s gone around it. I think that we’ve done a huge amount of work to deliver this model safely." Despite these assurances, caution is advised regarding granting GPT-6 Astra access to sensitive accounts in the near future.