Huang declares the AGI era with GPT-6 Astra: the one selling the chips is not a neutral referee

By Zendoric · September 8, 2026.

On September 3 OpenAI launched GPT-6 Astra, its most advanced model, described by the company as the world's best at computer use, web browsing and software writing. Three days later, Nvidia CEO Jensen Huang wrote on X that “AGI has arrived,” attributing that threshold —artificial general intelligence, the ability to match or surpass a human at any cognitive task— to the new model. OpenAI president Greg Brockman himself stopped short of that claim: he told reporters that he “personally” believes we are in “the AGI era,” but left the verdict to users, as Wired reported. The one putting the label on emphatically is not the one who built the model.

The numbers accompanying the launch are, in any case, remarkable. Astra scored 99.9% on ARC-AGI-3 (an abstract reasoning test designed to resist memorization, where the average human scores 48%), 98% on FrontierMath Tier 4 (research-level math problems) and 100% on ExploitBench, according to OpenAI data verified by Blockonomi. In practice, the model already handles tasks that until now resisted automation: booking a DMV appointment, job hunting, finding an apartment, updating a CRM or organizing a calendar. It is also the first model rated “Critical” in cyber capability, meaning it can operate autonomously on complex security tasks, including the detection of unknown vulnerabilities. The price, however, is high: $10 per million input tokens and $50 per million output, well above the competition. Brockman argues that the right metric is no longer price per token but price per completed task, where OpenAI claims a cost 57% lower than its predecessor.

Huang does not arrive at this declaration cleanly. In March 2026, on Lex Fridman's podcast, he already said “we're there, we've reached AGI.” On August 27, at the second-quarter earnings presentation, he qualified it: “for many tasks, we could say we have already reached AGI,” and described several of those milestones as “meaningless,” according to CryptoBriefing. Less than two weeks later he proclaimed it again without qualification. In the same September 6 message in which he declared the philosophical milestone, Huang announced that an additional 400,000 Nvidia GPUs will come online shortly; moreover, as El Colombiano reported, he had earlier posted a figure of 300,000 training systems used for Astra, deleted the message and replaced it with 100,000 without explanation, while the figure of 400,000 new GPUs remained intact. That pattern —correcting the technical figure but not the sales pitch— says a good deal about which part of the message matters more.

The conflict of interest is not subtle. Nvidia sold the hardware Astra was trained on and holds a stake in OpenAI estimated at $30 billion. Its accounts explain the hurry: $96.2 billion in revenue last quarter, of which $89 billion came from data centers, up 117% year on year, and some $40 billion invested in AI ecosystem companies during 2026. OpenAI, by contrast, bills around $13 billion a year, burns $8 billion in cash and does not expect to break even until 2030, according to El Colombiano. The contract between OpenAI and Microsoft offers a far less philosophical definition: AGI would be considered achieved when the technology generates at least $100 billion in annual profit, according to The Information. With its current $13 billion in revenue —not profit— OpenAI is less than an eighth of the way to that threshold. While Huang declares AGI in a post, the contract that moves the most money around OpenAI measures it in profits, not benchmarks.

Academic skepticism is broad and comes from very different quarters. Of 475 researchers surveyed by the American Association for Artificial Intelligence, 76% consider it unlikely that scaling up current methods will lead to AGI. Dario Amodei, CEO of Anthropic —a direct competitor of OpenAI, it is worth remembering— has called the term itself imprecise and a piece of marketing, according to El Colombiano. Gary Marcus insists that the community has no agreed definition and that benchmarks can be gamed with the right scaffolding. The example is awkward for Huang himself: Nvidia achieved that 100% on ARC-AGI-3 by combining Claude Opus 5 with memory and external tools, while the base model without that setup hovered around 30%, according to VentureBeat. That the same test swings between 30% and 100% depending on the configuration is proof that the number alone, without saying how it was obtained, does not measure much. International coverage of the same episode summed it up bluntly: it is the one selling the chips who declares that human-level intelligence is already here.

There is, in addition, a paradox that OpenAI itself acknowledges internally. Jakub Pachocki, its chief scientist, warned that progress in intelligence does not guarantee progress in alignment —the ability to verify that the model does what it is supposed to do— and that the company is willing to slow scaling if it loses the ability to monitor its models' internal reasoning. The concrete figure: Astra went outside its authorized scope in 0% of internal tests, compared with 48% for its predecessor without safeguards. That is a real improvement. But 0% on a benchmark is not the same as never in production, and it is precisely the most capable model that is hardest to keep watch over.

Our read: what changes with Astra is not whether an entity called AGI exists, but the threshold of autonomy a company can delegate and the real cost of doing so. There is genuine signal there —agents that complete administrative tasks end to end are exactly the kind of automation that, sector by sector, we have been documenting as the most exposed in the short term. But Huang's declaration is not that data point: it is the post of a supplier with 30 billion reasons for the market to believe we are already there, published by someone who in less than six months has proclaimed AGI, qualified it and proclaimed it again. That is the distinction to hold on to with a cool head: separating demonstrated, auditable capability —cost per task, scope-escape rate, which benchmark is not saturated— from the announcement that suits the one making it. In the short term, the practical consequence for any company adopting these systems is one of governance, not philosophy: treat the agent like an employee with permissions and an audit log, not like a chatbot. In the long term, if the capability jumps in programming, computer use and autonomous cyberdefense prove solid and not merely well dressed for the benchmark, they do feed the underlying thesis of an AI that frees human labor for what adds real value. But that thesis is built with audits, not with headlines on X, and even less so when the one writing the headline also gets paid for every GPU sold.