Notably, this is the first OpenAI model to rely significantly on other AI models to supervise its own training process.
The release is significant because Astra is designed to act as an autonomous agent, performing complex professional tasks directly within software rather than merely offering suggestions.
In demonstrations, the model formatted legal contracts, designed printed circuit boards, and contributed to scientific research in mathematics and physics.
However, Astra is also the first model to reach a "critical" cybersecurity risk threshold under OpenAI’s safety framework, meaning it has the potential to find and exploit system vulnerabilities without human guidance.
This capability led the company to delay the release to implement additional safety testing and "stay in bounds" of user intentions.
Astra will initially be available to a select group of organizations through a "Daybreak Access" program before rolling out to ChatGPT Plus, business subscribers, and API developers in the coming days.
While the model shows advanced reasoning, OpenAI executives acknowledged that Astra is more difficult for humans to monitor than previous versions.
To address this, the company is prioritizing research into "chain-of-thought monitoring," a method of tracking the step-by-step reasoning process the AI uses to reach a conclusion, to ensure the system remains under human oversight as its autonomy grows.