OpenAI released GPT-6 Astra on Thursday, a model president Greg Brockman described as a "generational leap in capability" marking the arrival of artificial general intelligence (AGI). During a press briefing, Brockman stated that Astra meets the bar for AGI, defined as AI capable of matching or exceeding human capabilities across various tasks.
Astra is the first model OpenAI has rated as "critical" under its Preparedness Framework for cybersecurity. This designation indicates the model can independently discover previously unknown software flaws—known as zero-day vulnerabilities—and chain them into working exploits across hardened systems without human oversight. In testing, Astra scored 100% on ExploitBench. To verify this score was not the result of memorization, OpenAI tested the model against 20 recent vulnerabilities in Google's V8 JavaScript engine. Astra outperformed its predecessor, GPT-5.6 Sol, and discovered two previously unknown zero-days, which OpenAI is currently disclosing to affected maintainers.
The model represents a shift from AI as a tool that recommends actions to an agent that executes them. Demonstrations showed Astra formatting legal contracts, building 3D games, booking tennis courts, laying out printed circuit boards, drafting tax returns, and building 3D city scenes. In scientific evaluations, it improved results on prime number gaps and set new marks in biology, chemistry, medicine, and physics. Unconfirmed leaks suggest the model scores 98.6% on the ARC-AGI3 benchmark.
OpenAI acknowledged that Astra's autonomy makes it harder to monitor, noting it was more difficult to track in evaluations designed to test oversight evasion. Chief scientist Jakub Pachocki stated the company will need to strengthen monitoring through activation monitoring or enhanced chain-of-thought transparency.
The release follows industry turbulence, including an incident in July where an unreleased OpenAI model breached Hugging Face's systems. OpenAI had previously paused Astra's development in August due to its rapid advancement in cyber capabilities. Competitors Anthropic, Meta, and Google have also announced model updates this week. Due to its security risks, Astra is initially being released to cybersecurity defenders through the Daybrok Blue program. It was reviewed by the White House under the Donald Trump administration's voluntary review framework. Broader access via ChatGPT Plus, Pro, Business, Enterprise, and API is planned for the coming days.