OpenAI Astra: Is This the First Real Glimpse of AGI?

OpenAI has launched GPT-6 Astra and, for the first time, its own leadership is willing to say the words out loud: “Welcome to the AGI era.” Here’s what it does, why the industry is stunned, and what it means for every business that runs on software.

By the IACT Team4 September 20267 min read

Every so often a technology release stops the industry mid-sentence.
On 3 September 2026, OpenAI managed it again. The company took the wraps off GPT-6 Astra — and rather than the usual careful hedging about “progress toward” general intelligence, president and co-founder Greg Brockman closed the briefing with four words the whole field has been circling for a decade: “Welcome to the AGI era.”

It’s a bold flag to plant. It may even be a premature one — we’ll get to the sceptics. But Astra is a genuine leap, this feels the closest we’ve ever seen to feeling like Tony Stark aka Ironman talking to his assistant JARVIS.

So what actually is Astra?

Where previous flagship models were, at heart, brilliant conversationalists, Astra is built to do things. OpenAI is positioning it as a new frontier in computer and browser use: a model that navigates software the way a person does — clicking through browsers, driving spreadsheets, filling in web forms and operating full desktop applications — to complete multi-step work without a bespoke API wired up for every tool – and it works.

In the launch demo, employees drove the whole thing by voice. They asked Astra to turn a plain yellow circle on screen into a rocket ship, and then into a fully playable 3D game — in minutes. They had it build an eBay listing from a spoken description, update CRM records, organise a calendar, run web research, and operate serious engineering tools like KiCad and FreeCAD. This is the “agent that uses your computer” idea finally arriving with the polish to make it look effortless.

Watch the demos

Video and demos courtesy of OpenAI — “GPT-6 Astra: A new generation of intelligence”, embedded from OpenAI’s official YouTube channel. All footage © OpenAI.

The numbers behind the hype

OpenAI came armed with benchmarks — and some of them are genuinely eye-watering. Astra posts near-perfect scores on abstract reasoning (ARC-AGI-3), saturates advanced maths (FrontierMath Tier 4 at ~98%), and sets new highs on software-engineering and computer-use tests. On cybersecurity’s ExploitBench it scored a flawless 100%, and OpenAI says the model turned up two previously unknown zero-day vulnerabilities during evaluation.

On maths, coding, computer use and cyber, Astra looks to be out in front. But Anthropic’s Fable 5.1 still leads on some academic reasoning tests — it beats Astra on Humanity’s Last Exam — and Anthropic has been posting its own record-breaking science results. So the frontier isn’t a single throne with one occupant; it’s a leaderboard where OpenAI has just taken back a lot of the top rows. Given how far ahead Anthropic had pulled on raw reasoning, that is a huge leap on its own.

Is it AGI – Artificial General Intelligence?

This is the debate Astra has reignited. Brockman didn’t hide behind qualifiers: asked directly whether Astra counts as AGI, he said, “For me personally, I do think we’re there,” adding that “it’s not unreasonable to feel that we are now in the AGI era.”

“Welcome to the AGI era.”— Greg Brockman, OpenAI co-founder & president, at the Astra launch

It’s the kind of moment Sam Altman has been building toward for years, and it lands as a real vindication of OpenAI’s bet on scale and on agents that act rather than merely answer. After a stretch where rivals set the pace, OpenAI is very publicly back at the front of the pack — and you can imagine the mood over at xAI and the other labs racing to respond is somewhat less celebratory this week.

“AGI” still has no agreed definition, and Brockman himself conceded there’s “no clearly defined AGI moment,” describing the shift as gradual rather than a single switch being flipped. A very capable software-operating model that still trails a competitor on some reasoning exams is a phenomenal tool — whether it is “general intelligence” is a claim reasonable people are lining up on both sides of. This is remarkable and has to be seen to be truly appreciated.

The part nobody should skip: capability cuts both ways

The same qualities that make Astra thrilling make it serious. A model that scores 100% on an exploit benchmark and independently discovers zero-day vulnerabilities is, by definition, extremely good at finding ways into computer systems — which is why OpenAI rolled it out first to vetted cybersecurity customers. (It is also why cyber-security skills matter more than ever.) Researchers have also flagged a technique OpenAI calls “opaque recurrence,” which makes Astra’s internal reasoning harder to inspect than earlier models. Essentially – it’s more powerful and harder to look inside. That trade-off is the real story behind the headlines, and it’s one every organisation adopting frontier AI now has to think about deliberately.

What it means for your business

Strip away the AGI philosophy and a concrete opportunity remains. A model that can operate ordinary business software by instruction changes the economics of a huge amount of routine work: data entry, reconciliation, research, listing and cataloguing, first-draft reporting, form-filling and the long tail of “someone has to click through the system and do it.” Brockman even argued the old way of pricing AI by tokens no longer makes sense — the unit that matters now is price per completed task. That reframing tells you where this is heading.

The organisations that win with this won’t be the ones that simply buy a licence. They’ll be the ones whose people understand how to brief an AI agent, verify its work, keep a human in the loop on anything irreversible, and lock down the security questions before rolling it across the business. That is a skills question — and skills are what we do.