GPT-6 Astra Launches as OpenAI Declares 'AGI Era' Has Arrived
View original source →OpenAI launched GPT-6 Astra on September 3 — a flagship model built to operate software autonomously and stay on complex jobs for hours or days — while Greg Brockman simultaneously declared that the AGI era has arrived. The rollout is limited and deliberate, because Astra is the first commercially released model to trigger OpenAI's highest internal safety classification.
Key Points:
• OSWorld 2.0 benchmark: Astra scored 72.6%, averaging 40 minutes per task — compared to the previous flagship Sol at 65.7% and 75 minutes per task.
• ARC Prize score: 62.7% with a neutral setup, and approximately 99.9% when OpenAI's built-in reasoning harness was active — demonstrating that agent performance depends heavily on the surrounding system architecture.
• Astra is the first OpenAI model to reach the company's 'Critical' cybersecurity threshold, indicating meaningful ability to assist sophisticated cyberattacks. It found two previously unknown software vulnerabilities during pre-launch testing.
• OpenAI notes that Astra's written reasoning became harder for human monitors to follow, which partially explains the limited rollout.
• Access is rolling out to limited partners first, then paid ChatGPT plans, API, Azure, and Bedrock over coming days.
• Early tester reports: Latent Space burned 20 billion tokens and called Astra 'an AI engineer you can hire for less than $6 per hour.' Ethan Mollick noted less drift on long jobs — Astra is better at not resurrecting discarded ideas.
The Critical threshold triggers mandatory extra safeguards — this is the first commercially released model where access rules are as consequential as the performance jump. The 'harder to monitor' written reasoning finding is the most significant safety signal in the launch notes — it means traditional interpretability checks are becoming less effective as models become more capable.
Why It Matters: The Critical cybersecurity threshold triggers mandatory safeguards that will affect deployment in regulated industries. The gap between capability and human oversight is widening, defining the governance challenge of the next phase of AI.