Jensen Huang says AGI is here. The evidence says the debate is far from over. 

When Nvidia CEO Jensen Huang declared that “AGI has arrived” following OpenAI’s launch of GPT-6 Astra, it sounded like one of the biggest milestones in computing history. If true, humanity would have crossed the threshold into artificial general intelligence: machines capable of matching or surpassing human intelligence across almost any cognitive task. 

There is just one problem. Almost nobody agrees on what AGI actually is. 

GPT-6 Astra is unquestionably OpenAI’s most capable model yet. The company says it can autonomously use computers, conduct complex research, write production-grade software, perform advanced scientific analysis, and even identify previously unknown cybersecurity vulnerabilities. Some of those capabilities have pushed Astra into OpenAI’s highest cybersecurity risk category, forcing the company to impose additional safeguards before wider deployment. 

Those are remarkable achievements. But remarkable is not necessarily general. 

Also read: OpenAI’s Astra raises a bigger question: Who watches AI? 

A moving goalpost 

The biggest obstacle to Huang’s claim is that AGI has never had a universally accepted definition. 

Historically, OpenAI described AGI as systems that outperform humans at “most economically valuable work.” Independent analysts note that Astra has not demonstrated that benchmark publicly, despite impressive gains in autonomy, coding, and reasoning. 

Even OpenAI CEO Sam Altman has repeatedly acknowledged that AGI remains a loosely defined concept, making absolute declarations difficult to verify. That leaves the industry arguing over a finish line that nobody has officially drawn. 

The critics are not buying it 

AI researcher Gary Marcus has become one of the sharpest critics of Huang’s announcement, dismissing the claim as lacking both evidence and clear definitions. His argument is straightforward: extraordinary claims require measurable criteria, not celebratory headlines. 

Others make a subtler distinction. DeepMind researchers have proposed a five-level framework for AGI, ranging from emerging and competent to expert, virtuoso, and superhuman. Under that framework, today’s leading language models could represent important progress without necessarily qualifying as full human-level intelligence across all domains. 

In other words, Astra may be closer to AGI than its predecessors without actually arriving there. 

Capability versus reliability 

OpenAI’s own rollout reflects that tension. Astra has demonstrated extraordinary performance in cybersecurity testing, including discovering zero-day vulnerabilities under controlled conditions. At the same time, the company warns that increasingly capable models can conceal aspects of their reasoning, complicating oversight and making alignment harder. 

Thus, if developers still need stronger shutdown mechanisms, restricted deployments, and additional monitoring before broader release, declaring victory over AGI begins to look premature. 

OpenAI’s chief scientist, Jakub Pachocki, has even argued that society may need to slow AI development because increasingly autonomous systems pose risks ranging from cyberattacks to deception. 

Why Huang might still be saying it 

Huang’s optimism is hardly surprising. Nvidia’s chips power the infrastructure behind frontier AI models, including Astra, making every leap in capability a validation of the company’s central role in the AI economy. Huang’s optimism is thus akin to a baker egging on fat kids to gorge on goodies to their heart’s content. 

His broader argument is that practical usefulness matters more than philosophical definitions. If AI is writing software, conducting research, and completing valuable work independently, obsessing over labels may be missing the bigger picture. 

Perhaps so. But history has a habit of punishing industries that mistake technological acceleration for completed revolutions. GPT-6 Astra undeniably pushes AI into new territory. It can reason better, work longer, and act more autonomously than previous models. Yet the industry’s own leaders cannot agree on whether that equals AGI, and the companies building these systems are simultaneously celebrating their power while warning about unprecedented risks. 

The safer conclusion may be this: AGI has become less of a destination than a contested narrative. Astra may have changed the conversation, but it has not ended the argument. And if AGI is indeed here, there needs to be incontrovertible evidence of that before bombastic claims can be dropped into conversations.