OpenAI says we’ve entered the AGI era. Is it true?
Has the AGI era finally begun? OpenAI says GPT 6 Astra could mark the start, but experts say benchmark scores alone do not prove human-level intelligence.
AI has crossed another line, at least in the way Silicon Valley is talking about it.
OpenAI’s launch of GPT-6 Astra has sparked a fresh debate over whether the world has entered the AGI era. Artificial general intelligence, aka AGI, typically refers to AI that can either match or outperform humans across a wide range of valuable tasks.
Astra appears to be pushing that idea further. But calling it AGI is a much bigger claim. Here's what you need to know!
Why OpenAI thinks the AGI era has begun
OpenAI President Greg Brockman reportedly said it is “not unreasonable” to feel that the AGI era has begun with Astra. He pointed to the model’s ability to operate computers, move through spreadsheets, fill forms and navigate web pages at very high speed.
OpenAI describes GPT-6 Astra as its most capable and aligned model so far. The model is designed to deliver state-of-the-art performance in software engineering, web browsing, computer use, cybersecurity, science and professional work.
Its reported benchmark scores are also striking. Astra scored 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench.
Nvidia CEO Jensen Huang added fuel to the debate by saying “AGI has arrived” after Astra’s release. He also highlighted Nvidia’s hardware. Astra was trained on more than 100,000 Grace Blackwell NVLink72 systems. Now, around 400,000 GPUs are expected to be deployed soon.
Why researchers are still saying “not so fast”
Not everyone is ready to put the AGI label on Astra. AI researcher Gary Marcus called Astra “pretty impressive” and acknowledged it as a genuine advance. At the same time, he argued that success on ARC-AGI does not, by itself, prove AGI.
His concern is whether Astra’s capabilities remain robust when it faces open-ended, real-world tasks rather than carefully defined benchmarks. Google DeepMind CEO Demis Hassabis is also preparing for a world in which AGI may be closer than many expect.
In a recent proposal, Hassabis said AGI could arrive within the next few years and argued that increasingly capable frontier models need stronger, more rigorous and regularly updated evaluations. He has proposed a US-based standards body to assess advanced AI systems before deployment.
Marc Andreessen, meanwhile, has taken the most direct position, arguing that AGI is already here. The disagreement comes down partly to definition.
A model can perform exceptionally well across benchmarks without necessarily possessing the broad judgement, adaptability, social understanding and real-world reliability associated with human intelligence.
So, has AGI actually arrived?
The safest answer is: not in any settled or universally accepted sense. Astra appears to represent a major jump in AI capability, particularly for long-running digital tasks. It may also make the idea of an “AGI era” feel more plausible to companies and users.
But impressive benchmark scores do not automatically prove human-level general intelligence. Benchmarks test specific abilities under specific conditions. Human intelligence involves judgement, context, responsibility, social understanding and adaptability in the real world.
OpenAI may be right that AI has entered a new phase. Whether that phase deserves the AGI label is another question entirely.