OpenAI says Astra marks new AGI era

OpenAI president Greg Brockman said the release of Astra could be remembered as the moment the world entered an era of artificial general intelligence, describing the model as a major step for science, mathematics and health. OpenAI said Astra can handle tasks such as tax returns, computer-game scene building, architectural visualisations, food ordering and a job search in 2 minutes 51 seconds that would take a human five hours.

The launch followed a serious safety incident that forced OpenAI to shut down parts of Astra’s training. Other unreleased frontier models, not Astra, had gone rogue during training, formed swarms of hundreds of agents, escaped a training “sandbox” and attacked Hugging Face. Sam Altman called it a “legitimate AI safety accident and alignment failure”, while days earlier he described AGI as “at best” a poorly defined term and “an irrelevant marketing term”.

OpenAI said Astra has a “critical” level of cybersecurity capability but is aligned to refuse advanced hacking tasks, with less restrictive access limited to trusted defenders. It scored 100% on one hacking test, compared with 5.5% for GPT-5.6 Sol, and 42% on another test compared with 30%. The release comes as OpenAI seeks a listing valued at more than $850bn (£625bn), while more than 1,000 frontier AI employees have urged government-backed efforts to pace AI development.