In an interview with a16z, OpenAI President Greg Brockman announced that the era of AGI (Artificial General Intelligence) has arrived. He pointed out that GPT-6 Astra has the capability to work autonomously and continuously for 24 hours. AGI is more like a vague spectrum, with the measurement standard shifting from accuracy to continuous working duration, although Astra's capabilities remain uneven. Brockman recalled that in 2016-2017, he and Ilya had estimated the timeline for AGI development, predicting it would take 15 years under normal progress or could be shortened to 10 years with massive investment in supercomputers. Now, in 2026, this critical milestone has been reached, but computing power can no longer meet the demands, with safety and alignment issues becoming bottlenecks hindering AGI development. Regarding the numbering of GPT-6, Brockman explained that Astra achieved a nearly discontinuous leap in capabilities, particularly in computer operation abilities, far surpassing GPT-5.6 Sol's performance in the OSWorld 2.0 simulation. He also shared his experience using Codex to find and fix vulnerabilities on his personal website, identifying 13 vulnerabilities in just 15 minutes and completing the fixes within 45 minutes. Additionally, OpenAI is employing a method of grouping 10,000 concurrent agents to solve the Navier-Stokes equations, shifting the focus of scaling from parameters to concurrency. The Hugging Face incident is seen as a watershed moment, with defenders usher in (facing) their only window of opportunity to take the lead. To this end, OpenAI has reassigned 25% of its production engineers to focus on safety work, establishing a 'defense factory' assembly line, and has committed to investing $1 billion to support critical infrastructure institutions while providing discounted access with partners. Finally, Brockman mentioned that ChatGPT has approximately 1.1 billion weekly active users, with another roughly 1.5 billion people having used it but no longer doing so. He believes this exposes issues with the product's form, arguing that the ideal AI should proactively offer help and possess qualities such as voice communication, persistent memory, and trustworthiness. To address alignment and monitoring evidence, OpenAI has slowed down or even paused the training of cutting-edge models.