On 2026-09-08, OpenAI announced its next-generation AI model, "GPT-6 Astra." OpenAI stated that Astra delivers state-of-the-art performance in fields such as computer operation, web browsing, software engineering, and cybersecurity.
The model features improved alignment between the understanding of user intent and actual behavior, enabling it to autonomously handle routine tasks such as filling out online forms, updating CRM records, and organizing calendars. For specialized operations, it also supports frontend QA (Quality Assurance) for websites, software installation and testing, and troubleshooting. Simulations using OSWorld 2.0 showed that processing time per task was reduced by approximately 47% compared to the previous generation, GPT-5.6 Sol.
In the field of cybersecurity, the model recorded a score of 100% on ExploitBench, which evaluates the ability to execute programs by exploiting vulnerabilities. Internal evaluations confirmed capabilities for arbitrary code execution in browsers and privilege escalation in operating systems using known vulnerabilities. In response, OpenAI stated that it has strengthened safety safeguards to prevent the model from executing tasks outside of intended scopes.
GPT-6 Astra began rolling out to select organizations today and is expected to become available via ChatGPT Plus, Pro, Business, Enterprise, and the OpenAI API within the next few days. It will also be provided through Microsoft Azure and AWS Bedrock.
Source: