“Welcome to the AGI era,” Greg Brockman told a Thursday press briefing, making an unusually direct claim for OpenAI’s newly released GPT-6 Astra. Brockman, OpenAI’s president and cofounder, said the model could come to be seen as the point at which artificial general intelligence, or AGI, arrived.
“If we fast-forward a couple of years, and we look back and say, "When was it, really, that AGI was created?" I think it's going to be about this time, and I think it might be about this model,” Brockman said. He added: “For me personally, I do think we're there … I think it's not unreasonable to feel that we are now in the AGI era.”
OpenAI released GPT-6 Astra on September 3, 2026. The company called it a “generational leap in capability” in cybersecurity, professional work, software engineering, science and computer use. Yet Brockman’s declaration turns on a definition that remains OpenAI’s own: the company once defined AGI as an automated system that can perform all economically valuable work as well as or better than humans.
Whether Astra can meet that test across the full range of economically valuable work is unknown. The model’s performance has not been independently verified outside OpenAI’s testing, and the broader AI research community’s view of Brockman’s classification is also unknown. The claim therefore marks a significant statement by one of the field’s most prominent companies, rather than a settled conclusion.
Astra’s defining feature is what OpenAI calls “computer use”. The model is designed to work directly inside software, rather than simply recommend what a person should do next. It aims to navigate a computer as a human would, including across systems without step-by-step human guidance.
Brockman said he had been thinking about an AI using a computer as a human does since he first started at OpenAI in 2015. The company says Astra can potentially find and exploit previously unknown vulnerabilities across well-protected systems without that detailed human direction. Such capabilities sit at the centre of both OpenAI’s case for the model’s advance and its concerns over releasing it.
Astra is the first OpenAI model designated as reaching the company’s “critical” cybersecurity threshold under its preparedness framework. OpenAI previously slowed the model’s release to add safety testing after determining that its cyber capabilities could reach that threshold. The specific concerns that prompted the delay, and how they were resolved, are unknown.
OpenAI said it worked with the government on testing before release. Mia Glaese, who leads the company’s safety processes, referred to a new misalignment monitoring approach that includes “24/7 escalation and rapid response” for potential concerns. Under that system, researchers are notified within 30 minutes.
The company has also pointed to benchmark results. On ARC-AGI-3, Astra scored 99.9% using a souped-up harness and 66% using ARC-AGI-3’s standard harness. It scored 100% on ExploitBench, a cybersecurity challenge. These results illustrate strong performance on the selected tests, though they do not by themselves establish that Astra can carry out all economically valuable work at or above human level.
Astra was built on OpenAI’s largest-ever training run, using more than 100,000 GPUs at its Stargate site in Texas. It is also OpenAI’s first model to use other models in a significant role in supervising its training. The company is initially limiting release to a limited set of enterprise customers, including participants in its cybersecurity-focused Daybreak program.
OpenAI said Astra would become available to Plus, Pro and Enterprise users “in the coming days”, as well as through the OpenAI API and AWS. The phased rollout reflects the tension behind Brockman’s announcement: OpenAI is presenting Astra as evidence of an AGI transition while treating its ability to operate computers and identify vulnerabilities as a capability requiring additional safeguards.
Comments (0)