OpenAI’s new generation of artificial intelligence no longer just responds: it can directly perform a sequence of actions on a computer.
OpenAI introduced GPT-6 Astra on September 3, 2026. The model is designed to handle tasks involving multiple steps, particularly in a browser or professional software. For example, it can fill out forms, update data in management software, organize a calendar, or conduct online searches.
The OpenAI name alongside a stylized brain made of electronic circuitry.
Credit: ChatGPT, CC0, via Wikimedia Commons.
The difference from a conventional text assistant is tangible. Astra can use a computer’s interface to carry out certain operations itself when it has the necessary tools and permissions. It can also install and test software, identify problems visible on the screen, or check whether a website is working properly.
This is an integrated version of OpenAI’s assistant. Solutions that come close to this behavior have of course already existed for several months, but they required installing third-party tools and using models that were not specifically designed for these tasks.
OpenAI also announced progress in programming. In Codex, Astra can work on long projects while keeping notes across multiple context windows. This allows the model to recall a requirement, a test result, or the reason for a previous failure without relying solely on a conversation summary.
The announced improvements also concern speed. In a simulation based on the OSWorld 2.0 test, Astra achieved a score of 72.6% in about 40 minutes per task. GPT-5.6 Sol reached 65.7% in about 75 minutes in the same simulations.
The model can also produce documents, spreadsheets, and presentations by following existing templates. OpenAI says it selects useful information more selectively instead of using all the available context. For users, the goal is to be able to delegate a coherent sequence of operations rather than guide every small step individually.
This increased autonomy also raises security questions. Astra is the first model that OpenAI has classified as “Critical” for its cybersecurity capabilities under the company’s own evaluation framework. With suitable tools, it can identify previously unknown vulnerabilities and develop methods to exploit them. The company says it has strengthened the restrictions and monitoring surrounding these uses.
OpenAI also says it has worked on respecting limits set by the user. During an internal test without the protections used in production, GPT-5.6 Sol exceeded the permitted limit in 48% of the cases studied. Astra did so in none of the cases in that evaluation. However, this was a test designed by OpenAI, not a guarantee valid for every real-world situation.
Deployment is beginning with a limited number of organizations. OpenAI then plans to make Astra available to ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as through its programming interface. Access is expected to expand gradually, making it possible to observe its behavior on real-world tasks that are far more varied than the tests presented at launch.