There is something strange about human intelligence.
We are physically weaker than countless things around us. Yet we built cities, crossed oceans, reached the Moon and created machines that can now reason with us.
The advantage was never strength.
Humans are unusual because we can look at a situation, understand it, choose between different paths, change our mind when something goes wrong, and act on that decision. Raw power is everywhere in nature. Judgment is rare.
That is also why AGI has fascinated the AI world for so long. AGI is not simply an AI that knows more facts. The bigger idea is a machine that can understand a goal, reason about it, make decisions, use tools, learn from what happens, change direction and keep working across very different kinds of problems.
Then, on September 3, OpenAI released GPT-6 Astra.
At the launch briefing, OpenAI president Greg Brockman ended with four words:
“Welcome to the AGI era.”
Maybe history will decide that sentence was early.
But Astra makes the AGI argument much harder to laugh away.
Something changed beyond the benchmark
Astra can obviously write an email or explain Python.
That part is boring now.
What makes this model interesting is what can happen after the answer would normally end.
You can give it a website and ask it to test the site. It can open the browser, click through the product, inspect what happens and continue working.
Give it a codebase and it can investigate the problem, edit files, run the application, test the result and keep fixing things.
Give it files, spreadsheets or an existing presentation and OpenAI says it can produce finished professional work while following the format already there.
The computer-use numbers also moved sharply. On OSWorld 2.0, Astra scored 72.6%, compared with 65.7% for GPT-5.6 Sol. More interestingly, OpenAI reports that the average simulated task fell from about 75 minutes to 40 minutes.
Then there is the stranger part.
Astra is OpenAI’s first model to reach its Critical cybersecurity capability level. OpenAI says research versions could find previously unknown vulnerabilities and develop exploit chains against hardened systems.
So yes, this is a very powerful model.
But the biggest change for normal users is simpler:
Stop giving Astra questions. Start giving it jobs.
That is where this guide begins.
Read more
