OpenAI unveils GPT-6 Astra and Nvidia certifies the 'AGI era': the problem is that nobody agrees on what that means

By IndexBox · September 8, 2026.

OpenAI has unveiled GPT-6 Astra, its new flagship model, with improvements in computer use, programming, scientific research and office work. The rollout began with enterprise customers who have access to the "Daybreak" program and is now extending to ChatGPT Plus, Pro, Business and Enterprise users, as well as to the OpenAI API and to Amazon Web Services (AWS), Amazon's cloud. After the launch, Jensen Huang, co-founder and chief executive of Nvidia, wrote on social media that "the era of artificial general intelligence (AGI) has arrived," congratulated OpenAI and specified that the model was trained on a large number of Nvidia Grace Blackwell NVLink72 systems —server clusters with ultra-high-speed interconnects, designed to train models at scale—. Huang added that the leap from ChatGPT to o1 and from there to Astra took place over four years, and that 400,000 additional GPUs are about to come online.

According to OpenAI, Astra can handle tedious administrative tasks —filling out online forms, updating customer records, organizing calendars, doing research— and generate documents, spreadsheets and presentations. The company presents it as its best model to date for software engineering, with strengthened cybersecurity capabilities. Chief executive Sam Altman highlighted that one of its most striking abilities is spotting problems the user had not raised. OpenAI president Greg Brockman told reporters ahead of the launch that the model marks a turning point and that, in his view, "it's reasonable to feel" the industry has entered the AGI era; that sentence is best read for what it is, the opinion of an executive at the very company launching the product, not an outside verdict.

Here the underlying problem appears: not even Huang and OpenAI agree on what the AGI they claim to have reached actually is. OpenAI defines it as highly autonomous systems capable of outperforming humans at most economically valuable work. Huang and Nvidia, by contrast, define it as a system capable of passing professional certifications and standardized exams with top-tier scores across multiple fields. These are very different bars: passing exams is measurable and has been demonstrated for some time now across several frontier models; "outperforming humans at most valuable work" is a far broader claim, and much harder to verify with an announcement and a social media post. That Nvidia is also the largest supplier of the chips needed to train the next model does not invalidate Huang's observation, but it does explain why he has every incentive to inflate the term: the sooner AGI is declared, the more urgent it seems to keep buying GPUs.

It is no coincidence that, among the coverage of this same launch, Investing.com ran the headline that GPT-6 Astra "is a real leap, but not the AGI moment OpenAI is selling," while other outlets simply repeated Huang's line with almost no qualification. That divergence is, in itself, the most honest news: when the specialist press splits between "AGI has arrived" and "this is a real improvement dressed up as a historic milestone," the prudent move is to assume the second reading until independent benchmarks exist —third-party evaluations, not company press releases— that place Astra against its rivals. The material consulted does not include those comparative figures, and the sector's own track record of "imminent AGI" announcements advises waiting to see them before accepting the claim.

The launch was also marked by an uncomfortable question: according to the article, Altman had to answer questions about a security breach in an earlier test model exposed on Hugging Face, the platform where AI models are hosted and shared, and he detailed Astra's strengthened security standards in response. The article offers no further details on the scope of that incident, so we treat it as what it is —an open question, not a settled fact— but it is worth keeping in view: the more autonomous a model becomes in handling accounts, forms and customer records, the greater the risk surface if something goes wrong.

Our reading: this announcement is a good illustration of why at Zendoric we insist on separating demonstrated capability from the marketing narrative. That a model improves substantially in programming, research and office tasks is real news and consistent with the trajectory of recent years; that this amounts to "AGI has arrived" is an interpretation, and one that suits whoever sells the chips as much as whoever sells the model. In the long run we remain convinced that systems of this kind, if deployed sensibly, will end up freeing human time from the administrative and routine tasks that today consume much of the working day, and that abundant compute (those 400,000 GPUs Huang mentions) is a necessary condition for getting there. But the road to that abundance is not shortened by declaring AGI in a social media post: it is shortened with independently audited models, with security incidents explained transparently, and with benchmarks anyone can reproduce. Until that arrives, Huang's statement is, at the very least, premature.