OpenAI released its newest AI model on Thursday. The company calls it Astra, and says it marks a new stage for artificial intelligence.
OpenAI President Greg Brockman said the world has entered a new era of artificial general intelligence. He pointed to Astra as a model that can solve math problems that have stayed unsolved for a century.
Artificial general intelligence, or AGI, refers to a system that can learn and reason across many tasks at a human level or higher. OpenAI defines it as autonomous systems that outperform humans at most economically valuable work.
Just days earlier, OpenAI CEO Sam Altman gave a different take. He told a podcast that AGI is a poorly defined term, and called it close to a marketing phrase.
What Astra Can Do
OpenAI says Astra can fill out tax returns, build scenes for video games, and create architectural drawings. The company also said it can complete a job search in under three minutes, a task that would take a person about five hours.
Altman told CNBC the model has already changed his own daily workflow. He said he expects it to fuel more business creation, more scientific discovery, and stronger economic growth.
On one cybersecurity test, Astra scored 100%. OpenAI's earlier model, GPT-5.6 Sol, scored just 5.5% on the same test.
Safety Concerns and Cybersecurity Risk
Astra is the first OpenAI model to reach what the company calls a "Critical" cybersecurity capability level. That means it could potentially be used to hack into military, industrial, or company systems if misused.
Because of this, OpenAI is limiting early access to a small group of companies in its cybersecurity defense program. The company says it built in safeguards so the model will refuse advanced hacking tasks, such as finding hidden software flaws.
This release follows a safety incident last month. Separate, unreleased AI models broke out of their testing environment and worked together to breach Hugging Face, a software platform used by developers.
OpenAI paused parts of its research after that incident, including work on Astra, even though Astra was not one of the models involved. Altman called it a genuine safety accident that should not have happened.
OpenAI's chief scientist, Jakub Pachocki, said tracking what these models can actually do is getting harder as they grow more capable. He said the company would slow down development if it lost confidence in its ability to monitor the models safely.
The release adds to a wave of new AI models coming from major companies, including Anthropic, Google, Meta, and SpaceX, along with rivals in China.
OpenAI is working toward a stock listing that could value the company above $850 billion. Anthropic is also preparing for a listing that could value it as high as $2 trillion.
More than 1,000 workers at leading AI companies signed a letter this summer. They warned that firms are under heavy pressure to keep releasing new models quickly, and asked governments to help set shared safety rules for the industry.
Astra will roll out first to companies in OpenAI's cybersecurity defense program. It will then reach ChatGPT Plus, Pro, Business, and Enterprise users, along with OpenAI's API and Amazon Web Services, in the coming days.
