AGI Made Clear › The declaration
OpenAI’s president says the AGI era has begun
“Welcome to the AGI era.” Said at a product launch, about a product, by a company whose own chief executive stopped defending the word a year ago.
At the GPT‑6 Astra launch on 3 September 2026, OpenAI president Greg Brockman said it was “not unreasonable to feel that we are now in the AGI era”, and that looking back in a couple of years “I think it’s going to be about this time, and I think it might be about this model”.
He also conceded, in the same briefing, that AGI is a “gray, fuzzy thing”, and described the label as a mission concept rather than a contractual threshold. Both things were said in the same room on the same afternoon, and the second is the more interesting.
Why the claim is defensible on its own terms
OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”. That is an economic test, not a cognitive one - it can in principle be satisfied by a system that resembles human thinking not at all.
Measured against that bar, the claim has evidence behind it. On GDPval, blind expert graders rated frontier output at least as good as a human expert’s 84.9% of the time in April 2026, against 47.6% seven months earlier. Judged economically, something has clearly changed.
And why it is not the end of the argument
Three weeks before the launch, Nvidia’s Jensen Huang had already said AGI arrived, on a bar of his own - whether a system can build a billion-dollar business. Three months before that, Google DeepMind’s Demis Hassabis put it at “2030, plus or minus a year”, on a bar requiring all the cognitive capabilities of the human brain. Nobody has changed their mind. They are marking different exams.
Sam Altman, asked on CNBC in August 2025 for his personal definition of AGI, answered: “I have many, which is why I think it’s not a super useful term.” His own essays stopped using it as a threshold from mid-2025. The era was declared by his president, in a vocabulary he had already set down.
What would make it more than a launch line
The measurements that have not moved are the ones worth watching. In July 2026, Epoch AI reported no evidence that the ability to learn on the fly has improved at all across model generations. In the same month Anthropic reported its own new model hallucinating slightly more than its predecessor. And the gap between what a frontier model does once and what it does reliably remains roughly eight to one.