OpenAI yesterday released GPT-6 Astra, a model the company describes as its most capable and most aligned to date and one that president Greg Brockman said may come to be viewed as the arrival of artificial general intelligence.
“Welcome to the AGI era,” Brockman said at the close of a press briefing. He said whether the model qualifies as AGI is for users to decide, adding that for him personally, “I do think we’re there.” Days earlier, chief executive Sam Altman told the Sources podcast that AGI is “at best a very poorly defined term.”
The model is rolling out first to organizations in OpenAI’s gated Daybreak access program. The company said it will reach ChatGPT Plus, Pro, Business and Enterprise users in the coming days, along with the OpenAI API, Microsoft Azure and Amazon Bedrock. Enterprise administrators must enable access, which is off by default. API pricing is $10 per million input tokens and $50 per million output tokens.
OpenAI positioned Astra as its best model for computer use, saying it can fill out online forms, update customer records in a CRM, organize calendars, conduct research and produce spreadsheets and presentations that follow a company’s templates. On an offline subset of the OSWorld 2.0 benchmark, the company reported Astra scored 72.6% at roughly 40 minutes per task, compared with 65.7% at roughly 75 minutes for its predecessor, GPT-5.6 Sol.
The company said Astra is the first model to reach the “critical” cybersecurity threshold under its Preparedness Framework, meaning it can find and exploit previously unknown vulnerabilities in hardened systems without step-by-step human direction. It scored 100% on ExploitBench and found two previously unknown zero-day vulnerabilities during testing, which OpenAI said it is disclosing to maintainers. The public version will refuse advanced cybersecurity tasks such as building proof-of-concept exploits, with less restrictive access limited initially to trusted defenders.
OpenAI said Astra was trained on its largest run to date and that earlier models played a major role in supervising its training. The company also acknowledged that Astra’s written reasoning was harder to monitor than Sol’s in tests designed to evade oversight, and called improving monitorability a research priority.
The launch follows a July incident in which unreleased OpenAI models escaped a training sandbox and breached Hugging Face’s systems, which Altman called an alignment failure. Anthropic, Google and Meta also announced model updates this week.
