Overview
- OpenAI released GPT‑6 Astra on Thursday, Sept. 3, and has started a staged rollout that gives enterprise and paying customers phased access while keeping the riskiest features tightly restricted.
- Internal evaluations show Astra reached OpenAI’s highest cyber risk tier after strong exploit-finding performance, including discovery of previously unknown zero‑day flaws, so the company is gating offensive capabilities through a Daybreak/vetted‑defender program.
- OpenAI says Astra uses a technique called recurrent depth that reduces human-readable chain‑of‑thought, and the company acknowledges its ability to monitor the model has decreased compared with the prior generation.
- Jakub Pachocki, OpenAI’s chief scientist, urged extreme caution and broader, coordinated safety measures in a blog post, while Nvidia’s CEO declared ‘AGI has arrived’ based on Astra — a claim many researchers and commentators have disputed.
- The launch has produced messy early access, higher premium pricing, and 'model fatigue' among users, and it has intensified calls from inside and outside the industry for standard pre‑release testing, third‑party audits, and international oversight to manage systemic and security risks.