OpenAI releases GPT-6 Astra with controlled high-risk safeguards

OpenAI has launched GPT-6 Astra for selected enterprise customers, according to reports. The company claims major advances in reasoning, coding, vision and long-horizon agent work, and says advanced access has been restricted because of the model's critical cyber capability.

Why it matters: as frontier models gain autonomy, the security burden of releasing them rises. A model capable of sophisticated cyber work is valuable to defenders and dangerous in the wrong hands, which is why gated access has become the industry's default for its strongest systems.

For organisations deploying such models, the recommended controls are strict: isolated execution environments, deny-by-default network access, short-lived credentials and human approval for consequential actions.

The performance claims come from the company and are not independent. Enterprises evaluating GPT-6 Astra should test it on their own tasks under their own conditions rather than relying on launch materials, and should review what the gated-access terms mean for their use cases.