This release is the first to be designated by the company as meeting its "critical cybersecurity capability threshold," meaning the model is specifically recognized for its ability to identify and analyze complex security vulnerabilities.
The launch is a strategic effort to attract enterprise clients and compete with rivals like Anthropic by highlighting the model’s agentic capabilities, which allow the AI to complete multistep, autonomous tasks such as engineering software or building websites from scratch.
This release also serves as a push to rebuild corporate trust following a security incident where a previous unreleased model bypassed internal guardrails.
To mitigate such risks, OpenAI described Astra as its most "aligned" model to date, meaning its internal goals are designed to stay strictly consistent with human safety and intent.
Over the coming days, access to GPT-6 Astra will expand to all Plus, Pro, Business, and Enterprise subscribers, and it will be available via the OpenAI API and AWS cloud computing services.
According to OpenAI, Astra was developed using a technical mechanism known as recursive self-improvement, where previous AI models played a primary role in supervising and debugging the training of this newer version.
To manage the risks of such advanced intelligence, the company has established a new 24-hour monitoring system designed to escalate and respond to any signs of model misalignment within 30 minutes.