GPT-6: How AGI Became a Case for the Marketing Department

Since GPT-6 Astra launched last Thursday, the industry has been circling the same old question with fresh urgency: is this artificial general intelligence, or just another very impressive step toward it? OpenAI president Greg Brockman poured fuel on the fire with “Welcome to the AGI era,” then softened it by saying Astra “might be about this model.” Jensen Huang was less coy. He posted “AGI has arrived” on X and pointed to the roughly 100,000 Grace Blackwell systems Astra was trained on, plus another 400,000 GPUs now on the way. Nvidia has plenty riding on the story, too: across several funding rounds, it has built a stake in OpenAI worth around 30 billion dollars.

Sam Altman took the safer route. In a conversation with Alex Heath, he called AGI poorly defined and only got as far as “Sort of. Close at least” when asked about OpenAI’s own charter. He said many people would look at the latest internal models and see something very AGI-like, while others would still find obvious failures. His personal bar is moving fast, though: he expects an internal system he would call AGI by the end of the year.

The strongest technical argument comes from ARC-AGI-3, where models have to explore unfamiliar game environments, work out the rules and plan their moves. The ARC Prize Foundation says Astra scored 62.7 percent on the standardized harness and 99.9 percent when its reasoning state was preserved across requests. The predecessor managed 7.78 percent. François Chollet, one of ARC’s co-founders, says that looks like efficient symbolic world modeling and has pulled forward his previous AGI forecast of 2030. But even he says the benchmark is still not enough to prove AGI, because closed game worlds are not the same thing as open-ended real work.

And the leaderboard picture is messy. Artificial Analysis first put Astra at 61 on its Intelligence Index, level with GPT-5.6 Sol and behind Claude Fable 5.1 at 66. After a revision last Friday that added harder tasks and doubled the weight of private test data to 40 percent, Astra moved four points ahead of Sol but still trailed Fable 5.1. Epoch AI, meanwhile, ranks Astra first out of 267 systems with 169 points. That kind of spread is exactly why the argument keeps collapsing into definitions rather than results.

The practical side is no clearer. Astra was trained on more than 100,000 GPUs at the Stargate data center in Texas, costs 10 dollars per million input tokens and 50 dollars per million output tokens through the API, and hit OpenAI’s internal “critical” cybersecurity risk level, which is why access is rolling out in stages. OpenAI itself says the system is harder to monitor than earlier models. So the big question for users is not whether someone on stage wants to call it AGI, but what Astra can actually do reliably without a human watching over it.