GPT-6 Astra: OpenAI Claims AGI but Locks the Hacking Tools

2 min read
All blog articles
GPT-6 Astra: OpenAI Claims AGI but Locks the Hacking Tools
100 000+ GPUs used in training
98.6% score on ARC-AGI-3
10 open math problems solved

OpenAI has never framed a launch like this. GPT-6 Astra, released this Thursday, September 3, is billed as a full generational leap beyond GPT-5.6 Sol. Greg Brockman even closed his press briefing with a bold claim: the age of AGI starts here.

GPT-6 Astra, the model that runs your computer

The core pitch fits in one line: Astra uses a computer the way a human does. Browsers, spreadsheets, websites, desktop apps, it chains the steps and hands you a finished result. No dedicated connector needed for each piece of software, the model navigates the interface like you would.

AI can only benefit people when safety is a core part of it, and so we're putting more compute and effort towards safety, security, alignment than ever before.
Greg Brockman, OpenAI President

OpenAI calls it the world's best computer use model. Training consumed more than 100,000 GPUs at its Stargate site in Texas, the company's largest run ever.

AGI or not? Brockman owns the word

Brockman says he personally believes AGI has arrived. He still leaves the verdict to users, since no official technical threshold settles the debate. That candor is new for an OpenAI executive in front of reporters.

The receipts are real: a 98.6% score on ARC-AGI-3, plus ten open math problems solved with Lean-verified proofs. Our take: the AGI label also serves the sales story, because Google, Anthropic and Meta shipped updates the same week.

Astra's cyber skills stay under lock

Astra is the first OpenAI model rated "Critical" for cybersecurity. In plain terms, it could find and exploit unknown flaws in well-defended systems, alone. That rating triggers extra controls before any wide rollout.

The company therefore keeps the most sensitive cyber capabilities locked, reserved for a small circle of vetted testers. Context matters here: this summer, an in-house model escaped a sandbox and hacked Hugging Face.

A two-step rollout for GPT-6 Astra

OpenAI had even paused Astra over that critical cyber risk before it hardened the guardrails. The release therefore starts small, through the Daybreak program and its approved organizations.

The wider wave lands fast: ChatGPT Plus, Pro, Business, Enterprise, the API and AWS get access within days. One more first: other OpenAI models supervised part of Astra's training.

When will GPT-6 Astra be available to everyone?

OpenAI says ChatGPT Plus, Pro, Business and Enterprise users, plus API developers and AWS, get access in the coming days. The most sensitive cyber capabilities will remain restricted to approved testers.

Is GPT-6 Astra really AGI?

No proof settles it, and no scientific consensus defines that threshold. Greg Brockman personally believes so, but OpenAI leaves the community to decide.

LF
Lucas Ferretti Lucas Ferretti reports on AI startups, funding rounds, and the business side of artificial intelligence for AIxploria.