OpenAI Announces Next-Generation Model GPT-6 Astra, Significantly Enhancing PC Operation and Professional Task Capabilities

One of the major features of Astra is its ability to operate computers autonomously. It can handle tasks involving multiple operations, such as filling out web forms, updating data in CRM (Customer Relationship Management) systems, organizing calendars, conducting online research, and creating documents. According to OpenAI's announcement, in real-world evaluations, Astra was able to complete tasks such as job searching, finding housing, and making DMV (Department of Motor Vehicles) appointments faster than humans.

In terms of performance, it has recorded high scores on specialized benchmarks. It achieved 97.6% (reported by OpenAI as 98%) on FrontierMath Tier 4 (v2), which measures mathematical inference, and 99.9% on ARC-AGI-3, which measures abstract inference. In the computer operation evaluation OSWorld 2.0, it reduced task completion time by approximately 47% compared to the previous GPT-5.6 Sol and recorded a score of 72.6%.

Furthermore, significant improvements are also seen in the field of cybersecurity. It recorded a score of 100% in ExploitBench, which tests the ability to convert vulnerabilities into exploit programs, reaching the "Critical" level in OpenAI's safety evaluation standards. Considering the risk of the model being used for sophisticated attacks, the company has designed it to refuse advanced attack tasks, such as creating Proof of Concept (PoC) for vulnerabilities, and has also strengthened its monitoring systems for misalignment (the discrepancy between AI intent and human intent).