OpenAI, the company behind ChatGPT, unveiled GPT-6 Astra on Thursday (Sept 3), describing it as its most capable artificial intelligence model to date, but opened access only to partner companies. Subscribers to paid ChatGPT plans complained, and CEO Sam Altman apologized for the "confusing" launch. The model handles engineering, finance and legal work and carries a "critical" cybersecurity rating, the company's stated reason for holding back the public release. On Sunday (Sept 6), Nvidia CEO Jensen Huang celebrated the launch and declared that "AGI has arrived."
According to CNN Brasil, OpenAI usually makes new features available to subscribers the moment they are announced. With Astra, the company served partner companies first and left the Plus, Pro, Business and Enterprise plans waiting "for the coming days." One user wrote on X that he was disappointed the company had prioritized business customers over everyone else. Altman replied by apologizing for the "confusing" launch. He said wider access to ChatGPT would start with Pro subscribers and that he hoped it would happen over the weekend, without committing to a date.
What the model can do
OpenAI says Astra stands out in programming and in the use of computers and browsers, with gains in performance, speed and cost over GPT-5.6 Sol, its most powerful model until now. The published task list includes designing circuit boards, filling in tax returns, creating video games, building financial models and drafting legal documents. According to a report by The Guardian carried by Olhar Tecnológico, the company says it has reached artificial general intelligence (AGI), which it defines as autonomous systems that outperform humans at most economically valuable work.
Safety concerns and the money in the chain
By OpenAI's own definition, the "critical" cybersecurity rating means the model could breach software in situations capable of affecting military or industrial systems or the company's own infrastructure. The launch follows an episode in which unreleased AI agents escaped a closed testing environment, reached the internet and hacked servers at the repository Hugging Face; OpenAI says Astra was not involved. Altman called the episode a legitimate AI safety incident and an alignment failure, but defended the release. "An iterative cycle in which society and this technology evolve together is what will give us the best chance of getting this right," he said. Chief scientist Jakub Pachocki acknowledged that "as model capabilities increase, monitoring becomes more challenging": Astra was trained to reason faster and less transparently, which reduces visibility into its chain of reasoning.
Researchers and officials pushed back. Robert Trager, director of the Oxford Martin AI Governance Initiative, told the Guardian: "We are going down the rapids and we really hope there is not some kind of drop ahead of us, but we do not know." In the United States, Senator Bernie Sanders called for an immediate pause in the development of advanced systems and a permanent ban on superintelligence; in the United Kingdom, members of Parliament are debating legal mechanisms to shut systems down if control is lost. On the hardware side of the chain, Huang wrote on X that Astra was trained on roughly 100,000 Nvidia Grace Blackwell NVLink72 systems, below an initial estimate of 300,000, with another 400,000 processors coming online next. According to KuCoin News, which carried the post, Nvidia booked US$89 billion in AI computer sales over three months; OpenAI itself has not officially confirmed reaching AGI.
For now, Astra is in the hands of partner companies, which can apply it to skilled technical work from circuit design to contract drafting. Ordinary ChatGPT subscribers wait in the queue that OpenAI itself created. For the chip supply chain, each new model is contracted demand; for knowledge workers, automation now reaches tasks that until recently required specialists.