OpenAI Says “AGI” Is Here — What Does That Actually Mean?

When OpenAI released GPT-6 Astra in September 2025, Nvidia CEO Jensen Huang declared that artificial general intelligence had "arrived." OpenAI president Greg Brockman echoed the sentiment, welcoming everyone to "the AGI era." But Gary Marcus, the psychologist, neuroscientist, and longtime AI critic, isn't buying it. His diagnosis is blunt: the term has been hollowed out into a marketing slogan, and the industry's own predictions about what AGI would mean for society are being quietly abandoned.

What Astra Actually Does

Astra was trained on more than 100,000 NVIDIA GPUs at the Stargate Data Center in Texas, making it the largest training run in OpenAI's history at roughly two and a half times the cost of its predecessor. Marcus acknowledges that Astra is "a bit better than the models before" and outperforms its competitor Fable on several benchmarks. But he points to a recurring pattern he calls "benchmark maxing" or "benchmaxing": these systems are trained on the benchmarks themselves, score impressively in controlled tests, and then "don't typically do as well in the real world."