GPT-6 Astra improves computer use and better respects its limits

OpenAI has introduced GPT-6 Astra, the new version of its most capable AI model, replacing GPT-5.6 Sol. The company presents it as a more powerful system that is, at the same time, better aligned with the instructions and boundaries set by users.

Among its most notable uses is handling a computer autonomously: filling out forms, updating records, browsing the web, or installing and testing software. According to OpenAI, it performs these tasks more accurately and in 47% less time than its predecessor. The company also describes it as its best model to date for software engineering, as well as for producing documents, spreadsheets and presentations that follow existing templates and styles.

Among other mathematical advances, the model helped improve two results on gaps between consecutive prime numbers, one of which had remained unchanged for more than 80 years.

In a cybersecurity test that measures the ability to turn known vulnerabilities into working exploits, Astra scored 100%, up from 78.5% for GPT-5.6 Sol, and even discovered two previously unknown vulnerabilities, which have already been reported to those responsible for fixing them. Because of this leap in capability, OpenAI has classified Astra at the highest risk level, "Critical," under its own internal safety framework—something that had not happened with any previous model from the company. As a result, the model has certain advanced offensive tasks restricted, though OpenAI plans to gradually expand access.

Regarding its behavior, the company states that Astra better respects the limits of action set by the user and the working environment: in a test designed to check whether a model acts beyond its authorized scope when given an impossible task, the previous model did so 48% of the time, compared with 0% for Astra.

The model is now available to a limited set of organizations and will roll out over the coming days to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API, Microsoft Azure and AWS Bedrock. Pro, Business and Enterprise users will also get access to a higher-tier version, GPT-6 Astra Pro.

Key points

• OpenAI introduces GPT-6 Astra, its most capable model to date.
• It handles a computer more accurately and in 47% less time than its predecessor.
• According to OpenAI, it's its best model to date for software engineering.
• It helped improve two math results, one unchanged for over 80 years.
• In a cybersecurity test, it went from 78.5% to 100% accuracy.
• OpenAI classifies it at the "Critical" risk level in cybersecurity.
• It discovered two unknown vulnerabilities, already reported.
• It better respects action limits: from 48% to 0% of out-of-scope cases.