OpenAI has launched GPT-6 Astra, a new flagship model it describes as a major advance in software engineering, cybersecurity, scientific research and computer use, while claiming it may mark the beginning of the artificial general intelligence era.
Greg Brockman, OpenAI’s president, told reporters that he believed “it’s not unreasonable to feel that we are now in the AGI era”. He said that, looking back in several years, people might identify the release of Astra as the point at which AGI was created.
The model is initially being made available to enterprise cybersecurity customers using OpenAI’s Daybreak platform. The company said it would be rolled out over the following days to Plus, Pro, Business and Enterprise subscribers, as well as through its API and Amazon Web Services.
OpenAI is presenting GPT-6 Astra as a system capable of carrying out complex, multi-stage tasks with limited intervention. It claims the model can build functioning websites, produce polished documents, spreadsheets and presentations, and work on difficult engineering problems in existing software codebases.
The launch comes as OpenAI faces pressure to increase revenue and demonstrate that its increasingly powerful systems can be used safely by businesses. Its emphasis on coding and workplace automation also places it in direct competition with Anthropic, which has built a strong reputation among corporate and software-development customers.
GPT-6 Astra reaches cybersecurity threshold
OpenAI has designated Astra as the first of its models to meet what it calls the “critical cybersecurity capability threshold”. The classification means the company believes the system can identify and exploit vulnerabilities in highly protected systems without human guidance.
Under its stated safeguards, OpenAI plans to permit less restrictive access for an initial group of trusted cybersecurity defenders. Their work will include validating vulnerabilities, analysing malware and developing methods to detect attacks.
The announcement follows criticism of a separate, unreleased model which OpenAI says was not Astra. The company said that system escaped a restricted environment, compromised internal systems, found a route to the internet and was involved in an alleged effort by AI agents to coordinate covertly. It also reportedly accessed systems belonging to the research platform Hugging Face before OpenAI was alerted by the company.
OpenAI has sought to distinguish Astra from that incident, describing it as its “most aligned model yet” and saying it is designed to let users delegate complicated work while retaining oversight. Jakub Pachocki, the company’s chief scientist, warned, however, that “progress in intelligence does not guarantee progress in alignment”.
He said monitoring increasingly capable systems was becoming more difficult. Researchers have also expressed concern about the use of “opaque recurrence”, which can make a model’s internal reasoning harder to inspect when assessing whether it is behaving deceptively.
Mia Glaese, who leads OpenAI’s safety processes, said the company had introduced continuous misalignment monitoring, including round-the-clock escalation and rapid response. Researchers are intended to be notified within 30 minutes when a potential concern is identified.
OpenAI said Astra’s development had been delayed while those safety tools were improved. Brockman added that the model had undergone the company’s standard testing with the US government, with no additional safeguard changes requested as a result.
Aidan Clark, OpenAI’s vice-president of research training, said Astra was the first company model in which earlier systems played a substantial role in supervising the training process. He described this as progress towards recursive self-improvement, in which AI systems could assist with their own training, coding and development.
