Safety overview: GPT-6 Astra
2026-09-04 · OpenAI
Safety Overview: GPT-6 Astra
Model Positioning
GPT-6 Astra is OpenAI's most capable broadly deployed model. The phrase "broadly deployed" is significant: it indicates the model has been made widely available through consumer products or APIs, rather than being restricted to controlled research settings or limited access programs. Achieving the highest capability rating within that deployment context makes GPT-6 Astra a notable milestone in OpenAI's model lineup.
The Preparedness Framework
The Preparedness Framework is OpenAI's internal system for systematically evaluating the risk potential of frontier AI models before and during deployment. It assesses models across several key risk categories:
- Cybersecurity: The model's ability to assist with or execute cyberattacks, exploit software vulnerabilities, and conduct offensive or defensive cyber operations
- CBRN (Chemical, Biological, Radiological, Nuclear): The model's potential to provide meaningful uplift toward the development or deployment of weapons capable of mass casualties
- Persuasion: The model's capacity to influence human beliefs, decisions, or behavior at scale
- Model Autonomy: The model's ability to take consequential, goal-directed actions without ongoing human supervision
Each category is evaluated on a four-tier capability scale: Low, Medium, High, and Critical. The Critical tier represents the highest level of assessed capability—the point at which a model's abilities in that domain are considered sufficient to pose meaningful real-world risk.
Critical-Level Cybersecurity Capability
The defining characteristic of GPT-6 Astra in this safety overview is its cybersecurity rating. It is the first broadly deployed model to reach the Critical level under the Preparedness Framework's cybersecurity evaluation.
A Critical designation in cybersecurity indicates that the model's capabilities in this domain have crossed a threshold associated with significant potential impact on computer systems, network infrastructure, or information security. Under the Preparedness Framework's design, a Critical rating in any category requires that additional safety and mitigation measures be in place before a model can be deployed or continue to operate. GPT-6 Astra's broad deployment alongside this rating reflects that such measures have been implemented.
Implications for AI Safety Practice
GPT-6 Astra's evaluation under the Preparedness Framework carries several implications for how the AI industry approaches frontier model development and deployment:
- Structured evaluation scales with capability: As models become more powerful, the need for systematic, quantifiable safety assessment frameworks becomes more pressing rather than less. GPT-6 Astra's Critical rating illustrates the value of having defined thresholds that trigger specific responses.
- Deployment decisions tied to capability thresholds: The Preparedness Framework creates a direct link between assessed capability levels and deployment requirements. Models at the Critical level face higher bars for release, which ties commercial decisions to safety outcomes.
- Transparency as a foundation for oversight: Publishing safety overviews enables researchers, policymakers, and the public to engage meaningfully with questions about what frontier AI systems can do and how risks are being managed.
Summary
GPT-6 Astra represents a concurrent advance in both model capability and safety evaluation practice. As OpenAI's most capable broadly deployed model and the first to reach Critical-level cybersecurity capability under the Preparedness Framework, it marks a point in frontier AI development where the stakes of rigorous, transparent safety assessment are higher than ever.