OpenAI Says AGI is Here. But Not Everyone is Buying It

Has artificial general intelligence, or AGI, finally arrived? OpenAI believes it has. 

The company’s latest model, GPT-6 Astra, has reignited the debate after president Greg Brockman said the industry had entered the AGI era. NVIDIA CEO Jensen Huang too echoed the enthusiasm, declaring that “AGI has arrived” following Astra’s release.

The claims reflect growing confidence around the capabilities of frontier AI models. But these are standalone claims. There’s no universal agreement on whether AGI has truly been achieved. The core issue is that there is still no agreed definition or objective test for AGI.

Huang’s endorsement is notable because NVIDIA supplies much of the computing infrastructure behind the current AI boom. He had already made a similar argument earlier this year, saying on the Lex Fridman Podcast that he believed AGI was here. 

Why OpenAI Believes Astra is Different

OpenAI describes Astra as its most intelligent and aligned model, with state-of-the-art performance in computer use, browsing, software engineering, cybersecurity, science, and professional work.

Astra also posted unusually strong results on several demanding evaluations. OpenAI reported a 98% score on FrontierMath Tier 4 evaluating research-level math problems, a 99.9% score on the reasoning benchmark ARC-AGI-3, and a 100% score on the cybersecurity benchmark ExploitBench. The company also said Astra has helped solve longstanding open problems in mathematics.

Interestingly, ARC Prize Foundation, which is working to advance open-source AGI, explicitly mentioned that saturating ARC-AGI-3 should not be interpreted as proof of AGI. 

“When we launched ARC-AGI-3, we made it clear that saturating the benchmark would not represent ‘proof of achieving AGI’. Therefore, while we believe Astra represents meaningful progress towards generalisation, we are not claiming that it is AGI,” Greg Kamradt, President of ARC Prize Foundation, wrote in a blog post 

The benchmark tests agents in unfamiliar environments, but those environments remain bounded, deterministic, and closed-ended. The real world is considerably more open-ended.

The argument for AGI therefore rests less on Astra simply being better at one task and more on its ability to operate across a growing range of intellectual and professional tasks. That broader generality is central to the AGI argument.

Arun Chandrasekaran, Distinguished Vice President Analyst at Gartner, tells AIM that Astra’s performance on ARC-AGI-3 provides evidence of a step-function improvement in agentic generalisation in unfamiliar but bounded environments.

ARC-AGI-3 tests how AI agents perform in unfamiliar, abstract, turn-based environments. They must explore, infer goals, understand how the environment works and plan their actions without explicit instructions. 

Astra scored 99.9% when operating through OpenAI’s provider adapter harness. Notably, under its standard, provider-neutral harness, Astra scored 62.7%.

“Benchmark saturation does not establish AGI. The available evidence does not demonstrate generality across fundamentally different contexts,” Chandrasekaran notes.

He says a declaration of AGI could influence investor expectations by signalling a step-change in OpenAI’s long-term market opportunity, but it would not by itself determine valuation. “Public-market investors would likely focus on whether the claimed capability translates into durable revenue growth, total addressable market, and a defensible competitive position.”

Charlie Dai, VP and Principal Analyst at research firm Forrester, agrees that Astra represents a major improvement in how AI agents execute tasks. It combines stronger reasoning, computer use, software engineering, and the ability to complete multi-step tasks, while delivering much better benchmark results than earlier frontier models.

“However, this represents broader capability rather than proven general intelligence, so AGI remains an interpretation rather than an established fact,” he adds. 

Definition of AGI

OpenAI itself has historically defined AGI as highly autonomous systems that outperform humans at most economically valuable work. But that definition leaves considerable room for interpretation because it does not specify exactly which capabilities must be demonstrated, how reliably they must work, or how performance should be measured.

“The jury’s still out on what AGI even means, though it usually comes down to a system that reasons across domains, acts with real autonomy, learns from situations it was never trained for, and matches human performance on tasks nobody built it to do,” says Jacob Joseph, VP of Data Science at customer retention platform CleverTap. 

Joseph adds that OpenAI’s definition is an economic threshold, not a cognitive one, and it tells you the AGI label is doing more work as a market signal than as an objective description. “The test isn’t whether a model can title itself AGI; it’s whether it holds up to unsupervised, accountable use in the wild over time,” he notes.

That ambiguity matters because current AI models can be extraordinarily capable in some areas while still displaying significant weaknesses in others. A system can outperform humans on difficult mathematics or coding problems without necessarily possessing the flexible, reliable, and open-ended intelligence associated with human intelligence.

“What remains unproven is whether success on benchmarks and professional workflows translates into human-level, domain-independent intelligence—a threshold for which no universally accepted definition or test exists,” says Dai. 

No Universal Agreement 

AI researcher Gary Marcus has strongly pushed back against Huang’s claim.

“Unfortunately, Huang gave no evidence and no definitions, which feels to me like an effort at a takeover of a scientific question by corporate fiat,” he wrote on his Substack. 

Marcus argued that Astra may be impressive in areas such as coding and mathematical reasoning, but AGI is much more far-reaching. In his view, human-level general intelligence involves the ability to generalise reliably from experience to very different situations, demonstrate resourcefulness and operate with a level of reliability that current large language model systems have not demonstrated.

Joseph says currently there is no shared standard for declaring AGI, making such claims difficult to independently verify.

“AGI has no equivalent of a system card. There’s no shared, testable standard the industry has agreed to be measured against, so a declaration like this functions less as a verification and more as a proposal for a standard, made after the fact and offered for everyone to accept retroactively.”

Joseph says learning that transfers to a genuinely novel situation is close to impossible to demonstrate through a benchmark score, since a benchmark is, by definition, a task someone already anticipated.

“That’s why an AGI claim is better assessed over a protracted period of time, as these systems are tested in increasingly open-ended, real-world settings. A benchmark shows what a model can do when it knows it’s being tested. The more interesting question is what it does when there’s no test to game and the consequences are real,” he notes.

What Now? 

OpenAI and Huang are effectively arguing that the practical threshold for AGI has been crossed because AI systems can now perform a sufficiently broad range of economically valuable intellectual work. 

“Jensen’s endorsement adds credibility because NVIDIA has a unique vantage point on frontier AI progress through its role as the infrastructure provider behind leading AI labs,” says Dai. “His support should be seen as informed validation of accelerating capability trends rather than definitive proof that AGI has been achieved,” he adds. 

For now, the most defensible conclusion is that GPT-6 Astra represents a major expansion of frontier AI capabilities. However, whether it signals the arrival of AGI is still up for debate.