Summary
Every major OpenAI release triggers a new wave of AGI declarations, and GPT-6 Astra is no exception — Jensen Huang tweeted that AGI has arrived, while OpenAI President Greg Brockman simultaneously described it as a “gray fuzzy thing” with no clear definition. In this video, Fahd Mirza (filming from Sydney) cuts through the marketing noise to examine what Astra actually delivered in practice and where the limits remain.
The hands-on demo is concrete: using Hermes agent, Mirza hands Astra a Creative Commons-licensed Phoenix 3D model from SketchFab — an 89-bone rigged mesh with wings, tail, spine, and feathers — and prompts it to build an interactive bone viewer with no hand-holding. Astra scaffolds the entire project in a single pass, correctly reads the bone hierarchy from file rather than guessing, loads all textures and materials properly, and delivers a working interactive simulation. Notably, the model proactively flags what it cannot verify — including its inability to confirm the 3D scene rendered correctly — before Mirza redirects it to completion. Human-in-the-loop oversight remains necessary.
On the safety front, Astra is OpenAI’s first model to cross the critical cybersecurity threshold under their own safety framework, meaning it can find unknown vulnerabilities and build working exploits. During pre-release testing it uncovered two previously unknown zero-day vulnerabilities in Chrome, which OpenAI is disclosing responsibly. Mirza’s overall verdict: real capability gains in coding and agentic tasks, but the price point puts it out of reach for the overwhelming majority of users worldwide.
📺 Source: Fahd Mirza · Published September 07, 2026
🏷️ Format: Opinion Editorial
