OpenAI has launched GPT-6 Astra, a new AI model with advanced capabilities, including enhanced cybersecurity features and improved performance metrics.
On September 3, OpenAI launched its latest AI model, GPT-6 Astra, beginning a phased rollout for select organizations via its Trusted Access program. This access will soon be available to ChatGPT Plus, Pro, Business, and Enterprise users, along with those utilizing OpenAI API and Amazon Bedrock.
— OpenAI (@OpenAI) September 3, 2026
Microsoft has also started providing the model through its Microsoft Foundry program on Azure, with plans to expand access for clients shortly.
According to OpenAI, GPT-6 Astra surpasses its predecessor, GPT-5.6 Sol, in various tests related to programming, cybersecurity, and scientific tasks, although some models from Anthropic showed better performance in specific benchmarks.
OpenAI noted that the comparative metrics were based on the highest scores achieved in controlled environments, indicating that performance may differ for user-accessible versions.
Enhanced Functionalities
GPT-6 Astra is designed to autonomously navigate applications and browsers, allowing it to perform tasks such as filling out forms, managing CRM records, searching for information, and analyzing data. In the OSWorld 2.0 offline test, Astra scored 72.6%, outperforming GPT-5.6 Sol’s 65.7%. It also completed tasks in latency simulations approximately 47% faster, taking around 40 minutes compared to 75 minutes for its predecessor.
In the Agents’ Last Exam, which evaluates performance in real-world applications, Astra achieved a score of 59.3%, while GPT-5.6 Sol scored 53.6% and Claude Opus 5 scored 55.5%.
Astra demonstrated impressive results in evaluating understanding of computer interfaces, scoring 92.7% in ScreenSpot-Pro compared to GPT-5.6 Sol’s 76.9%.
OpenAI has also introduced an experimental context retention feature in Codex, which allows Astra to maintain notes across context windows and retrieve information from previous interactions. This feature will soon be a standard offering for Astra.
Advancements in Programming and Science
In the Terminal-Bench 4.0 for programming tasks, Astra scored 57.9%, significantly higher than GPT-5.6 Sol's 37.3%. On scientific benchmarks, Astra achieved 64.6% on Terminal-Bench Science 0.1, compared to 22.4% for GPT-5.6 Sol.
Astra excelled in FrontierMath Tier 4, obtaining a score of 97.6%. Developers noted that Astra made significant improvements in prime number theory, reducing a longstanding gap between infinite prime pairs.
However, in Humanity’s Last Exam, Astra scored 57.2%, which was lower than Claude Fable 5.1's 65%.
Cybersecurity Capabilities
GPT-6 Astra is the first OpenAI model to reach a Critical level in cybersecurity capabilities. It scored 100% on ExploitBench, significantly outperforming GPT-5.6 Sol's 78.5%. In internal tests, Astra successfully executed arbitrary code in 39% of cases involving high-severity vulnerabilities, compared to 11.5% for GPT-5.6 Sol.
"Astra's ability to identify and develop exploits for zero-day vulnerabilities could help find and mitigate weaknesses but simultaneously necessitates stricter security measures," OpenAI stated.
While the model includes additional restrictions to prevent advanced cyber tasks, OpenAI plans to gradually broaden access through its Daybreak initiative.
To enhance monitoring, a classification system will identify potentially unauthorized behaviors and halt them automatically, although Astra's reasoning may be less transparent than that of its predecessor.
$1 Billion Initiative for Cybersecurity
In conjunction with Astra's launch, OpenAI announced a $1 billion initiative named Daybreak for Frontline Defenders, which includes subsidized access to models and training for organizations with limited cybersecurity resources. Initially focused on U.S.-based operators in critical infrastructure sectors, the program is expected to expand internationally.
"Defenders have a window of opportunity: it is narrowing, and AI needs to be leveraged to fill security gaps ahead of attackers," OpenAI emphasized.
Daybreak is already in use by around 2,000 approved organizations, offering varying levels of access based on the sensitivity of the tasks involved.
Access and Pricing Details
ChatGPT Pro, Business, and Enterprise users will have access to a distinct version of GPT-6 Astra Pro. OpenAI has not detailed how this version differs from the standard one. Developers can access the model through the OpenAI API, with a context window of 1.05 million tokens and specific pricing structures based on token usage.
Users should be aware that in August, OpenAI temporarily paused the scaling of new AI models and reinforcement learning training for future deployments.
