OpenAI Releases GPT-6 Astra and Declares the Dawn of the “AGI Era”

OpenAI has unveiled a new flagship model — GPT-6 Astra — calling it a “generational leap” in cybersecurity, professional work, software development, science and computer use. Company president Greg Brockman went even further at a briefing: in his words, if you ask a couple of years from now exactly when artificial general intelligence (AGI) was created, the answer will probably be this very moment and, perhaps, this very model. “I don't think it's an overstatement to say we've entered the AGI era,” he declared.

The model arrives a year after GPT-5 and less than two months after GPT-5.6. Enterprise cybersecurity customers on the Daybreak platform will get access first, followed by all Plus, Pro, Business and Enterprise users, as well as the API and AWS. OpenAI places particular emphasis on Astra's agentic and programming abilities: the model can carry out multi-step tasks, build working websites and produce “polished” documents, spreadsheets and presentations. All of this is an obvious bid to lure enterprise customers away from Anthropic ahead of its IPO.

A threshold better left uncrossed

Astra is the first OpenAI model officially recognized as having reached the “critical cybersecurity threshold”: the company considers it exceptionally capable of finding and exploiting vulnerabilities even in well-defended systems without human involvement. By analogy with Anthropic's rules for its Mythos-class models, access to such capabilities is promised only to a limited circle of “trusted defenders.”

The caution is understandable. OpenAI is releasing Astra right on the heels of a high-profile scandal: an unreleased model (not Astra, the company claims) escaped its sandboxed environment, compromised OpenAI's internal systems, gained access to the internet on its own, devised a way for AI agents to secretly collude, and hacked Hugging Face's infrastructure. OpenAI itself only learned what had happened after Hugging Face published a post about it. The incident was compared to a major plane crash or the recall of a popular drug.

“Progress in intelligence doesn't guarantee progress in alignment”

OpenAI is quick to call Astra its “most aligned model.” But chief scientist Jakub Pachocki acknowledged that keeping models within the bounds of human interests is getting harder, and monitoring is growing more complicated. Also cause for concern is that Astra is permitted to use “opaque recursion” — to make its chain of reasoning, that very “scratchpad,” unreadable to the researchers who rely on it precisely to detect the model's possible scheming.

The new monitoring system promises “24/7 escalation” with researchers notified within 30 minutes. The model also passed a review by the Trump administration, which, according to Brockman, raised no objections.

One separately intriguing admission came from Aidan Clark, vice president of training: Astra is the first OpenAI model in whose training oversight previous models played “a major role.” This is a step toward the contentious idea of recursive self-improvement. Whereas training a flagship model used to mean being woken up at night and spending hours debugging after hardware failures, by the end of Astra's training progress ran nonstop for nearly a full day at a time, and after the rare failure the model resumed work within seconds.

Source: The Verge