OpenAI launches Astra, its powerful (and controversial) new model
TechCrunch · 2026-09-03
OpenAI has launched Astra, a new AI model that it claims is its most powerful and controversial yet, boasting superior cybersecurity and software engineering capabilities. The company states Astra achieved a perfect score on ExploitBench and discovered two zero-day vulnerabilities in modified tests, positioning it as the "best model for software engineering to date" by outperforming models like OpenAI's Sol and Anthropic’s Fable in tasks like bug finding and code query answering. Astra also incorporates a reasoning technique called "opaque recurrence," which raises concerns among AI safety experts like Redwood CEO Buck Shlegeris and Zvi Mowshowitz, as it obscures the model's chain of thought, making monitoring its decisions more challenging. OpenAI chief scientist Jakub Pachocki acknowledged that increased model capabilities make monitorability harder, while co-founder Brockman stated that for him personally, Astra signifies the arrival of artificial general intelligence (AGI), though the contractual definition of AGI with Microsoft has been removed. Access to Astra's advanced cybersecurity features will be limited due to safety precautions, which include new techniques to make the model safer and restricting responses to prompts from "higher risk" accounts.