OpenAI Launches GPT-6 Astra Amid Debate Over AI Safety and Monitoring
OpenAI has unveiled Astra, its latest and most powerful AI model, claiming unmatched speed, accuracy, and safety across computer and browser tasks. While celebrated for its advanced cybersecurity and software engineering capabilities, Astra introduces controversy with its use of opaque recurrence, which challenges traditional model monitoring. President Greg Brockman also offered a personal take on whether Astra signifies the arrival of Artificial General Intelligence, following an evolution in its definition.
OpenAI has launched GPT-6 Astra, its latest flagship model for complex reasoning, coding and computer use, describing it as a major advance in the capabilities of AI systems. The model is initially rolling out to enterprises in OpenAI’s Trusted Access Program, with access through the API and Plus, Pro, Business and Enterprise plans expected to follow.
OpenAI has placed particular emphasis on Astra’s cybersecurity capabilities, saying internal evaluations found it could discover previously unknown vulnerabilities and develop exploit chains, prompting the company to classify its capabilities as potentially reaching a critical threshold under its Preparedness Framework.
The model’s capabilities have also intensified scrutiny over how increasingly powerful AI systems can be monitored and controlled. OpenAI says it has introduced stricter security controls, sandboxing, restricted network and tool access, enhanced monitoring and additional alignment safeguards for Astra, including monitoring of its chain of thought during agentic activity.
The company has also clarified that Astra was not involved in the Hugging Face incident, although lessons from that incident were incorporated into its safety approach, making it inaccurate to suggest that the breach directly triggered Astra’s development or launch.
Astra’s rollout highlights the growing tension between expanding AI capabilities and the difficulty of ensuring those systems remain predictable and controllable. OpenAI says its evaluations found Astra more likely than GPT-5.6 Sol to respect explicit safety and security restrictions, while acknowledging that more capable models can create greater risks when given access to tools, networks and complex environments.
As access expands beyond specialised testing programmes, the central question will be whether Astra’s safeguards can keep pace with its ability to perform increasingly consequential work.