Did OpenAI’s GPT-6 Launch Just Deliver AGI? Jensen Huang Says Yes

OpenAI has pulled the trigger on what it calls its most capable artificial intelligence system yet, and the reaction from one of Silicon Valley’s most powerful voices has turned a product release into something closer to a moment. The OpenAI GPT-6 launch introduced a model called Astra on September 3, 2026, and within days Nvidia CEO Jensen Huang was on X declaring that artificial general intelligence had arrived. Not everyone in the AI world agreed with him, and that disagreement is arguably as newsworthy as the model itself.

Key takeaways

• OpenAI launched GPT-6 Astra on September 3, 2026, describing it as its most capable and aligned model to date.
• Nvidia CEO Jensen Huang wrote “AGI has arrived” on X on September 6, tying the claim to Astra’s training scale.
• Astra trained on more than 100,000 Nvidia Grace Blackwell NVLink72 systems, with another 400,000 Nvidia GPUs set to come online.
• The model posted a 98% score on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and a perfect 100% on ExploitBench.
• OpenAI says Astra hit its internal “Critical” cybersecurity threshold, meaning it can locate and exploit unknown software flaws — a capability the company has restricted at launch.

Inside the OpenAI GPT-6 Launch: What Astra Can Actually Do

Astra is designed to do far more than answer questions. OpenAI is pitching it as a genuine computer-use agent, one that can navigate browsers, spreadsheets, websites, and desktop software much the way a human employee would, rather than requiring developers to wire up a separate integration for every app it touches. During a closed briefing covered by VentureBeat, OpenAI co-founder and president Greg Brockman framed the shift bluntly, telling the room that the industry has been “bottlenecked” by teams “painstakingly building” connectors into tools people already know how to use.

In practice, that means Astra can fill out forms, update CRM records, comb through web research, manipulate spreadsheets, analyze data in Python notebooks, and even operate engineering software. OpenAI is rolling the model out first to enterprise customers through its gated Daybreak access program, before extending it to ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as developers using the OpenAI API and cloud platforms including AWS Bedrock and Microsoft Azure.

Astra’s Cybersecurity Powers Come With Guardrails

The most sensitive part of the release involves cybersecurity. OpenAI says Astra has crossed the company’s internal “Critical” threshold for that domain, meaning it can identify previously unknown software vulnerabilities and work out how to exploit them without step-by-step human direction. That is a meaningful jump in autonomy, and it’s exactly why GPT-6 cybersecurity capabilities are drawing so much scrutiny. OpenAI has responded by adding extra safeguards and limiting access to some of these functions right out of the gate, a sign the company itself sees real risk in what it built.

Jensen Huang Declares “AGI Has Arrived”

Nvidia’s chief executive didn’t wait long to weigh in. On September 6, Huang posted on X: “GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations OpenAI team. 400K GPUs coming online next.” The post did double duty — celebrating OpenAI’s milestone while quietly reminding investors just how much Nvidia hardware sits underneath the industry’s most advanced models.

The benchmark numbers OpenAI is showcasing back up the ambition. Astra scored 98% on FrontierMath Tier 4, a notoriously difficult mathematics benchmark, and 99.9% on ARC-AGI-3, a test built specifically to probe whether a system can generalize to unfamiliar problems rather than lean on memorized patterns. It also posted a flawless 100% on ExploitBench, the security-focused benchmark tied directly to its vulnerability-hunting abilities.

Those figures matter for reasons beyond bragging rights. This is exactly the kind of moment that fuels the broader AGI debate 2026 has been circling for months — a debate about whether raw benchmark performance is proof of general intelligence, or simply proof that a system is very good at the specific tests it was given.

Why the AGI Debate Isn’t Settled

Even inside OpenAI, the AGI label came with hedges. Brockman told VentureBeat that “everyone has a different definition of AGI,” admitting the company once expected a clean, universally recognized threshold that simply never materialized. Pressed on whether Astra clears that bar, he went further than most executives would: “For me personally, I do think we’re there. I think there’s a pretty good argument for it.” He summed up OpenAI’s institutional position as, “it’s not unreasonable to feel that we are now in the AGI era” — notably more cautious phrasing than Huang’s flat declaration.

Outside the company, the pushback was sharper. AI critic Gary Marcus rejected Huang’s framing outright, arguing there is no universally accepted definition of AGI and no clear evidence that any current system has actually crossed whatever line separates narrow capability from general intelligence. That skepticism isn’t limited to Marcus. A wider argument has broken out over what exactly gets measured when a model posts a near-perfect ARC-AGI-3 score — the underlying neural network, or the surrounding scaffolding of memory, tools, and orchestration layered on top of it. Nvidia’s own research team faced a version of this question in August, when its agent architecture hit a perfect score on the same benchmark using a different underlying model, illustrating how much performance can come from the system wrapped around a model rather than the model alone.

Why this matters: until the industry agrees on what AGI actually means and how to measure it fairly, claims like Huang’s will keep functioning more as marketing signals than settled scientific conclusions — which makes it harder for regulators, enterprises, and the public to calibrate how seriously to take each new milestone.

What Astra’s Scale Means for Nvidia and the AI Chip Market

Whatever the AGI question ultimately resolves to, the hardware story behind Astra is concrete. Training the model required more than 100,000 Nvidia Grace Blackwell NVLink72 systems, and Huang says another 400,000 Nvidia GPUs are about to come online — a scale of deployment that underscores just how tightly frontier AI progress is now bound to chip supply.

Markets have already priced in some of that momentum. Nvidia shares are up 23.7% year-to-date, and according to TipRanks, analysts covering the stock carry a Strong Buy consensus built on 29 Buy ratings, with an average price target of $325.23 — implying roughly 41.2% upside from current levels. For a company whose fortunes rise and fall with AI infrastructure spending, a launch of this scale, followed by promises of hundreds of thousands more GPUs, functions as a demand signal as much as a technology announcement.

That dynamic cuts both ways. If Astra’s capabilities hold up under real-world enterprise use, the appetite for Nvidia AI GPUs is likely to keep climbing through the rest of 2026. But if the AGI framing proves overstated once independent testing and broader deployment catch up with the marketing, the gap between hype and delivered value could become the next flashpoint in an industry that has learned, repeatedly, how quickly sentiment can swing on a single benchmark controversy.

FAQ

When was OpenAI’s GPT-6 Astra launched?

OpenAI launched GPT-6 Astra on September 3, 2026.

What notable GPU infrastructure was used to train GPT-6 Astra?

Astra was trained on over 100,000 Nvidia Grace Blackwell NVLink72 systems.

What is Nvidia CEO Jensen Huang’s view on Astra’s significance?

Jensen Huang declared “AGI has arrived” after the GPT-6 Astra launch, praising OpenAI’s achievement and noting that 400,000 more Nvidia GPUs are coming online.

Is there consensus in the AI community that Astra represents true AGI?

No. Experts including Gary Marcus point out there is no universally accepted definition of AGI, and even OpenAI’s own leadership described the company’s position in more measured terms than Huang did.

Article produced with the assistance of artificial intelligence and reviewed by the editorial team.