OpenAI unleashes Astra, its most capable and controversial model yet

Details

OpenAI has released GPT-6 Astra, its most capable model to date, and President Greg Brockman has publicly said it may represent the beginning of the "AGI era" — meaning AI that outperforms humans at most economically valuable tasks by OpenAI's own definition. The launch follows a familiar Daybreak-first rollout sequence, then ChatGPT subscribers and API customers on AWS Bedrock and Azure. Fast Company's framing centers on controversy. Astra's cybersecurity abilities are the sharpest concern: the model scored 100% on ExploitBench and autonomously discovered two zero-day vulnerabilities during internal testing — capabilities that have significant dual-use implications. OpenAI rates Astra as "critical" under its own safety framework, the first model to reach that designation. The second safety concern is interpretability: Astra's reasoning is reportedly harder to monitor than predecessors, widening the gap between model output and auditable reasoning chains. For enterprise teams relying on standard output inspection for compliance and governance, that represents a new deployment risk category. The Fast Company article is thin on benchmark specifics — it relies primarily on OpenAI's announcement framing. Detailed performance data (98.6% ARC-AGI-3, 97.6% FrontierMath, 74.1% DeepSWE) is available from concurrent reporting elsewhere. The safety-first lens here is the distinct editorial contribution.