GPT-6 Astra came out on 3 September, and within days the conversation wasn't really about what it could do. It was about whether it was AGI, artificial general intelligence, the long-promised point where AI matches people across the board. The quickest way to understand that argument is to look at who said what. They don't agree.
Who said what
- Greg Brockman, OpenAI's president: "It's not unreasonable to feel that we are now in the AGI era".1
- Jensen Huang, Nvidia's chief executive, whose chips Astra was trained on, posting on X: "AGI has arrived."2
- Sam Altman, OpenAI's own chief executive, who has described AGI as "not a super useful term".3
So the president of the company that built it says maybe. The man who sold it the hardware says yes. The chief executive has said the word itself isn't much use. That isn't three people reading the same evidence differently. It's three people using the same word to mean different things.
AI Fit Session
Not sure where to start? Over 1 week and three 60-minute 1-to-1 sessions, we map your business, find the quick wins, and build a practical starting point specific to you.
AI Foundations
Eight weeks, one coaching session a week, seven recorded modules. By the end you'll have AI working across real areas of your business and a 90-day plan to keep building.
AI Growth
Already using AI but want to keep developing it? Each month we work on real tasks from your business together, refining what's in place and building what's next.
The definition hardly anyone quotes
OpenAI does have a written definition. Its charter describes AGI as "highly autonomous systems that outperform humans at most economically valuable work".4 That's a test about work, not about how clever a model looks in a demo. And it's a high bar: most valuable work, done better than people, largely on its own.
Measured against that, Astra is impressive and still clearly short. It needs a paid plan, a person setting the task, and in most cases a person checking the result. That's a very capable tool. It isn't a replacement for most of the work people do.
What the people who run the test said
The loudest number was a 99.9% score on ARC-AGI-3, a benchmark built to probe general reasoning. Under the benchmark's standard setup, the same model scored 62.7%. ARC Prize, which runs the test, validated both results and then said plainly: "we are not claiming that it is AGI."5 When the people who built the AGI test won't use the word, a headline that does deserves some caution.
The lesson worth keeping
AGI isn't a finish line anyone has agreed on. It's a word that stretches to fit whoever is using it, whether that's to sell chips, raise money or play down a rival. None of that changes what matters for a business. The useful question is never "is this AGI?" It's "can this do a specific job for me, reliably, and how would I know?" That one can actually be answered, by testing it.
What is AGI?
AGI stands for artificial general intelligence. There is no single agreed definition. OpenAI's charter defines it as highly autonomous systems that outperform humans at most economically valuable work. Others use the term more loosely to mean AI that matches human ability across a wide range of tasks.
Is GPT-6 Astra AGI?
There is no consensus. OpenAI president Greg Brockman said it is not unreasonable to feel we are now in the AGI era, and Nvidia chief executive Jensen Huang said AGI has arrived. But ARC Prize, which runs one of the main benchmarks for general reasoning, said it is not claiming Astra is AGI, and the model still depends on people to set and check its work.
What has Sam Altman said about AGI?
OpenAI chief executive Sam Altman has described AGI as not a super useful term, reflecting his view that the word is poorly defined. That contrasts with OpenAI president Greg Brockman, who said GPT-6 Astra could reasonably be seen as the first model of the AGI era.
What did GPT-6 Astra score on ARC-AGI-3?
GPT-6 Astra scored 99.9% on ARC-AGI-3 using OpenAI's own test setup, which preserves the model's reasoning between steps, and 62.7% under the benchmark's standard setup. ARC Prize validated both results but stated it is not claiming the model is AGI.