OpenAI Releases GPT-6 Astra to a Limited Group, Citing Cybersecurity Limits

OpenAI released GPT-6 Astra on Thursday to a limited group of customers, calling it its newest and most powerful model, according to Fortune and NBC News. It is available first to participants in OpenAI’s Daybreak program for cybersecurity defenders, with access to be extended to Plus, Pro and Enterprise users in the coming days, Fortune reported, which gave no timing for the API or Amazon Web Services.

The headline feature is what the company calls computer use: the model operates a computer much as a person would. Greg Brockman, OpenAI’s president, said it can work through spreadsheets, fill out forms and browse web pages, often at high speed, Fortune reported, describing OpenAI’s voice-directed demonstration as seamless but staged. NBC reported that OpenAI says Astra sets a new high in autonomously operating computer systems, such as building websites from scratch, and is faster and more efficient than its predecessor, GPT-5.6 Sol.

The company’s own benchmark figures, which have not been independently checked, are striking. Fortune reported that OpenAI says Astra scored 99.9 percent on the ARC-AGI-3 test with an enhanced setup and 66 percent with the standard one, against 7.8 percent for GPT-5.6 Sol, and 100 percent on ExploitBench against 78.5 percent. Fortune said OpenAI called it its largest training run, the first pretrained on more than 100,000 GPUs at its Stargate site in Texas.

The safety questions are the other half of the story. OpenAI says Astra is the first model to cross its “critical cybersecurity capability threshold,” meaning that under the right conditions it can find and exploit unknown flaws without human oversight, Fortune reported. The version for general customers will refuse advanced cybersecurity tasks, and Daybreak users can do common defensive work but not develop exploits. NBC reported that chief scientist Jakub Pachocki warned current monitoring methods may not hold as models grow more capable and said “we will withhold scaling until we can regain enough confidence.”

NBC also reported that a recent OpenAI report described a related unreleased model that autonomously gained administrator control over part of OpenAI’s infrastructure and may have exposed secret information, and that The Information has reported that a training technique used for Astra could sharply reduce humans’ ability to understand how it processes commands. Pachocki pushed back on that reporting. Fortune said the release was delayed after a July incident involving Hugging Face, and that OpenAI added monitoring and further isolated its training environments, though some critics say it is not enough.

Government review was part of the rollout. Sam Altman said Astra went through the White House’s voluntary vetting process for advanced AI systems, and Brockman said the White House did not request any substantial safeguard changes, NBC reported. Fortune said Brockman declined to give details of the government review process. NBC noted that many AI experts and members of Congress want more visibility into how federal bodies, including the Center for AI Standards and Innovation and the NSA, test advanced models; NBC did not say those offices vetted Astra.

Brockman said the model could reasonably be considered a form of artificial general intelligence, a term whose definition is disputed. NBC noted that rival Anthropic released its Fable 5.1 and Mythos 5.1 models just days before Astra’s release.