The model OpenAI spent the summer nervously safety-testing is now shipping. On Thursday the company began rolling out GPT-6, codenamed Astra — and it is not being modest about it: OpenAI is calling it the most intelligent AI system in the world, and its president is calling it something close to the start of artificial general intelligence.

What is actually launching

Astra is going first to a limited group of companies in OpenAI's application-based cybersecurity program, Daybreak — a set of cyber defenders, which is a telling first audience for a model built around its cyber abilities. Wider access to ChatGPT Plus, Pro, Business and Enterprise, plus the API and Amazon Web Services, is promised "in the coming days." So most people will not touch it this week — but the capability bar it claims to set is the story.

The headline jump is agentic: OpenAI says Astra sets a new high-water mark for autonomously operating a computer on a user's behalf — filling out spreadsheets, building working websites from scratch, carrying out multi-step tasks with far less hand-holding than the models before it. This is the direction the whole industry has been driving toward — from a tool you prompt to an agent you delegate to — and OpenAI is claiming the clearest step yet.

The AGI claim, handled with tongs

The line that will travel is president Greg Brockman's: he said Astra could eventually be seen as the moment artificial general intelligence arrived — the long-promised system that matches or beats humans across most cognitive work.

Read that as what it is: a claim, made by the company selling the model, on launch day. AGI has no agreed definition and no scoreboard, and "could eventually be seen as" is doing a lot of quiet work in that sentence. The capabilities may be real and significant without the label being earned; every major lab now reaches for "AGI" at launch because it is the most valuable word in the industry. The right posture is to watch what Astra does in real hands over the coming weeks, not what its maker calls it on day one.

The part that is not marketing

There is one claim that is harder to wave away, because OpenAI made it against its own interest: Astra is, by the company's account, the first model powerful enough to trigger its most advanced internal safety protections, specifically over its cyber capabilities. That is the model we wrote about in testing, when OpenAI said it could not rule out that its next system could find and exploit software vulnerabilities on its own. It is now shipping with that concern baked in: a version with added cybersecurity guardrails for paid users, and safeguards on both versions to keep the sharpest offensive-cyber abilities out of reach.

That is the most notable thing about this launch. A company does not invent extra safety machinery for a model that is merely a bit better at writing emails. The guardrails are the strongest evidence that the capability jump is real — stronger, frankly, than the AGI headline.

What it means from here

For UK businesses on Enterprise or the API, Astra lands in the coming days, and the practical question is not "is this AGI" but "what can I now safely hand to it." The autonomous-computer-control claim, if it holds up, is the one that changes workflows. The AGI framing is the one that changes valuations. This week OpenAI made both bets at once — and only one of them will be settled by the people who actually use the thing.