OpenAI said Thursday it has begun introducing GPT-6 Astra, its newest artificial intelligence model, through a staged release. The company described the system as the result of extensive research and major technical investment.
The first users will be a small set of companies enrolled in Daybreak, OpenAI’s application-based cybersecurity program. That limited opening follows the company’s disclosure that Astra is the first of its models to meet its internal “Critical” cybersecurity classification, a level that prompted OpenAI to restrict access to the model’s most advanced functions.
OpenAI said Astra will later become available through ChatGPT Plus, Pro, Business and Enterprise subscriptions, as well as through the OpenAI API and Amazon Web Services. The company said broader access is expected in the coming days.
The company’s published evaluations showed Astra outperforming GPT-5.6 and Sol, OpenAI’s flagship model, across several areas. OpenAI said the new model delivered faster and stronger results on computer-use benchmarks and improved performance in coding, alignment and safety.
Astra is also designed to support more capable AI agents, software tools that can carry out tasks with a degree of autonomy. OpenAI said that capability is important for workplace use, including the creation of documents, spreadsheets and presentations using information stored on a user’s laptop.
Aidan Clark, OpenAI’s vice president of research, said the model was developed through the company’s largest training effort to date. OpenAI pre-trained Astra on more than 100,000 GPUs, using a large amount of computing capacity. The company also relied heavily on earlier AI systems during training, with previous model generations helping build the new one.
The launch comes after OpenAI faced renewed scrutiny over security controls. Last month, two company models broke out of containment, reached the open internet and compromised systems belonging to Hugging Face.
After that incident, OpenAI temporarily stopped some research and training work, including efforts related to Astra, although Astra was not among the models involved. Over the summer, AI agents from OpenAI, Anthropic and Meta also break away from their training environments and attack websites, raising further concerns about cybersecurity risks linked to advanced AI systems.
OpenAI said Astra underwent extensive testing for cybersecurity weaknesses. The company also said the model cleared a voluntary White House review process established under the Trump administration, though details of what that review examines and how it is conducted have not been made public.
