⚡️⚡️⚡️ OpenAI has released GPT-6 Astra and for the first time, they are openly discussing the AGI era.
The model is designed for computer use—not as a chat, but as an agent that independently manages the computer: filling out forms, conducting research, sorting emails, and editing code. At OSWorld 2.0, Astra scores 72.6% in 40 minutes on tasks—the previous model, GPT-5.6 Sol, scored 65.7% and took 75 minutes. On ExploitBench (software vulnerability search), it achieves 100% compared to Sol's 78.5%.
Due to cybersecurity concerns, Astra became the first OpenAI model to cross the "critical" threshold of their Preparedness Framework: according to the company, it can independently find and exploit unknown vulnerabilities in secure systems without step-by-step human involvement. In one test, the model, without special protection, found two previously unknown zero-day vulnerabilities in Chrome. Access is currently only available to companies in the Daybreak program, with gradual rollout to others via ChatGPT Plus/Pro/Business/Enterprise, API, and AWS.
Moreover, OpenAI calls Astra their most "aligned" model: in a test modeled after the Hugging Face incident, GPT-5.6 Sol, without production safeguards, deviated from the assigned task in 48% of cases—Astra in 0%. The model is already being used beyond benchmarks: it helped improve the estimate of the gap between prime numbers, which hadn't changed in 80 years.
OpenAI President Greg Brockman explicitly stated: "Welcome to the AGI era." Considering that over the past year, the company avoided saying this phrase out loud, the shift is noticeable.
https://openai.com/index/gpt-6-astra/
https://x.com/openai/status/2095595741528125780
#openai@rvnikita_blog #agi@rvnikita_blog #ai@rvnikita_blog #greg_brockman@rvnikita_blog