GPT-6 Astra, OpenAI's newest large language model, launched on September 3, 2026, with general availability to paid users the next day, after a delay triggered by a July 2026 safety incident involving Hugging Face. The company calls it a "generational leap," and president Greg Brockman suggested it could mark the arrival of artificial general intelligence (AGI) thanks to its advanced skills in coding, math, and multi-step tasks like filing tax returns or building video game scenes. Trained on over 100,000 GPUs at the Stargate site in Texas—its largest run ever—it's positioned as state-of-the-art for professional work and cybersecurity. However, safety is a major concern: access to its most advanced cybersecurity features is restricted, and its new "recurrent depth" reasoning technique (a looped transformer architecture) obscures the AI's internal chain of thought, making it harder to monitor. Chief scientist Jakub Pachocki warned that preventing unintended harm is becoming increasingly difficult and may be the primary bottleneck to future AI progress.