On Wednesday, OpenAI president Greg Brockman stood in front of a camera and told the world that GPT-6 Astra — the company’s newest model, rolling out this week to ChatGPT Plus, Pro, Business, and Enterprise users — represents “the start of AGI.” The pitch, repeated across the launch materials, is simple: “Anything you can do on a computer, Astra can do for you. Fast.”
That is the least interesting thing about this model.
A month before the launch, on August 1, OpenAI published a research post titled “Ten advances in mathematics and theoretical computer science.” The post reported that an internal version of Astra had produced new results on ten long-standing open problems, each accompanied by a machine-checkable proof in the Lean theorem prover. Not benchmark scores. Not “state-of-the-art on MMLU.” New mathematics — the kind of thing that, until now, required a human being with a decade of specialized training and, occasionally, a Fields Medal.
OpenAI buried that post. The launch is about computer use.
The Mathematics Post Nobody Read
The August 1 post is the most consequential thing OpenAI has published since the original GPT-4 technical report, and it received a fraction of the attention. Ten open problems. Machine-checkable proofs. That is not a demo. That is not a vibe. That is a verifiable claim about the extension of human knowledge, and it is either true or false in a way that no amount of marketing can fudge.
The Lean theorem prover matters here. Lean doesn’t take anyone’s word for anything. A proof in Lean is a proof, full stop. If Astra actually produced machine-checkable proofs for ten open problems, then we have crossed a line that most people — including most people in AI — did not expect to cross this decade.
And yet the launch narrative is about spreadsheets.
The 100% Number
Here is the other number from this week’s launch that deserves more attention than it’s getting: 100%. That is Astra’s score on ExploitBench, a benchmark that measures whether an AI system can turn known software vulnerabilities into working exploits. Without safeguards, Astra scored 100%. Its predecessor, GPT-5.6 Sol, scored 78.5%.
OpenAI’s own Preparedness Framework classifies Astra as reaching the “Critical” threshold in cybersecurity. The company is gating exploit-creation capabilities behind something called the “Daybreak program” — a name that sounds like it was chosen by someone who has watched too many spy movies and not enough security briefings.
The safeguards, as described, are “alignment training plus system safeguards like Codex Auto-review and misalignment monitoring in production.” That is a lot of words for “we sanded off the dangerous parts and we’re pretty sure it worked.”
Selling the Middle
Here is the thing that should make everyone uncomfortable, regardless of where they sit on the AI debate: OpenAI has built a system that can do original mathematics and turn any known vulnerability into a working exploit, and the company has chosen to market it as a butler.
“Anything you can do on a computer, Astra can do for you.” That is the pitch. Not “Astra extended the frontier of human knowledge.” Not “Astra can break your infrastructure.” The mundane middle.
One trader on a desk in lower Manhattan, watching the launch coverage between positions, put it this way: “They solved ten open problems and the headline is that it can book your flights. Either they don’t know what they have, or they know exactly and they’re trying to keep the conversation boring.”
The pricing tells the same story. $10 per million input tokens, $50 per million output tokens. That is not pricing for a research tool. That is pricing for a productivity suite. The mathematics is a footnote. The exploit capability is behind a door. The product is the middle.
The uncomfortable question is whether the middle is where the value actually is, or whether the middle is just where the liability is lowest. A model that can use your computer is a product. A model that can solve open problems in mathematics is a scientific instrument. A model that can turn any known vulnerability into a working exploit is a weapon. OpenAI has all three, and it is selling you the first one.
The launch of GPT-6 Astra is not the start of AGI. It is the start of something more mundane and more consequential: a company that has built instruments and weapons, and has decided that the safest thing to sell is a butler.