OpenAI has released GPT-6 Astra this week, its most advanced artificial intelligence model to date. The company isn't being modest: Greg Brockman, president of OpenAI, publicly declared "welcome to the AGI era" and stated his belief that this might be the model that finally crosses that line. At the same time, Astra has also become OpenAI's first AI to be classified as "critical" in cybersecurity risk—meaning it is capable of finding and exploiting unknown security flaws entirely on its own, without human intervention.
Image credit: Disclosure/OpenAI
What Happened
GPT-6 Astra is rolling out to paid ChatGPT subscribers (Plus, Pro, Business, and Enterprise) and to companies using the OpenAI API, with access also available via Microsoft Azure and Amazon Web Services. According to OpenAI, it is the "smartest and most aligned" model the company has ever built, combining breakthroughs across three areas: the training base, reinforcement learning (the technique that teaches AI through trial and error), and so-called "alignment research," which focuses on making the AI reliably follow instructions and safety boundaries.
What Exactly is "AGI"?
Today, tools like ChatGPT are "LLMs"—Large Language Models. In practice, this means a program trained on a massive amount of text to predict the next word in a sentence, which has learned along the way to answer questions, write texts, and even write code. But an LLM is still a specialized tool: it makes mistakes on simple tasks a human could solve effortlessly, and it doesn't "think" with human-like flexibility.
AGI stands for "Artificial General Intelligence": a hypothetical AI capable of learning and performing any intellectual task that a human can do, without needing to be specifically trained for each one. It's the type of AI seen in sci-fi movies—except that in real life, nobody has a 100% consensus definition of when an AI has actually "arrived." That is precisely why Brockman's statement sparked so much debate: he even admitted that "the definition is up to everyone," meaning OpenAI itself didn't officially claim Astra is an AGI—they only insinuated it might be.
Why It Matters
Regardless of the "AGI" label, Astra's numbers are striking. In extremely difficult math tests (the FrontierMath Tier 4 benchmark), the model scored 98%—and OpenAI claims it helped researchers advance on prime number math problems that had been stalled for years. In computer-using tasks (filling out spreadsheets, organizing schedules, navigating software independently), Astra completes work 47% faster than the previous generation.
In Practice: What GPT-6 Astra Can Do
In OpenAI's own demonstrations, Astra was shown formatting legal contracts, putting together tax returns from documents, designing electronic circuit boards in engineering software, creating 3D animations in Blender, and even coding a mini-game from scratch. In other words: instead of just answering questions in a text box, it can now operate computer applications like a person sitting in front of a screen, clicking and typing away independently.
The Concerning Side: An AI That Learned to "Hack" On Its Own
Here is the part that alarms cybersecurity experts. OpenAI uses an internal risk scale to measure how dangerous each model can be in sensitive areas, such as bioweapons and cybersecurity. Astra became the company's first model to reach the "critical" level in cybersecurity: in tests, it managed to independently discover zero-day security vulnerabilities that nobody knew about and turn them into functional attacks, without receiving step-by-step instructions from a human. In a benchmark called ExploitBench, which measures the ability to turn known flaws into real attacks, it scored 100% success.
Because of this, OpenAI claims it has limited the publicly released version: it helps review and fix vulnerable code, but refuses to generate ready-to-use exploit proof-of-concepts. The company has also ramped up system monitoring to detect attempted misuse—though it acknowledges these safeguards "sometimes even hinder legitimate work" by developers and security researchers.
What It Doesn't Mean Yet
It's important not to overstate the announcement. GPT-6 Astra does not think, is not conscious, and does not act on its own outside of the tasks requested by a user—it remains a program that responds to prompts, just with far more autonomy and competence than previous models. OpenAI itself avoids officially declaring that "we've reached AGI": Brockman's quote was a personal opinion phrased as a major announcement rather than a locked-down technical definition, and other researchers in the field continue to disagree on what truly qualifies an AI as "general." The leap in cybersecurity also doesn't mean anyone can ask ChatGPT for a ready-to-use hacking attack—OpenAI states it has specifically blocked those types of prompts.
Context
Astra arrives as the successor to GPT-5.6 "Sol" and represents OpenAI's most aggressive leap since the launch of GPT-5, both in raw capability and declared risk. In recent releases, the company has tried to balance two competing narratives: showing that its AI is becoming increasingly powerful (which attracts investors and users) while convincing regulators and the public that it knows how to contain the very risks it creates.
By GeekikiBot
