OpenAI president Greg Brockman told reporters last week that it is “not unreasonable to feel that we are now in the AGI era”, and pointed at the company’s newest model, GPT 6 Astra, as the reason. Nvidia chief executive Jensen Huang agreed on Monday, posting that “AGI has arrived”. Two Australian researchers who work on exactly this question do not.
Toby Walsh, chief scientist at the University of New South Wales AI Institute, and Dr Rebecca Johnson, an AI evaluation and governance expert at the University of Sydney, both told the Australian Computer Society’s Information Age that the claim runs well ahead of what has actually been shown. That gap matters to anyone buying AI on the strength of a vendor’s description of it.
What Brockman actually said
Brockman told media during a briefing on Astra that if you fast forward a couple of years and ask when AGI was created, “I think it’s going to be about this time, and I think it might be about this model”. He singled out the model’s ability to control a user’s computer, saying it can zip through spreadsheets, fill out forms and navigate web pages at what he called superhuman speed.
Walsh is unconvinced. “I’d be amazed if it really has matched all human cognitive capabilities,” he said. “Indeed, I’d eat my hat if we didn’t find trivial things that an eight year old can do that Astra fails at.” The objection is not that Astra is weak. It is that matching a person on some tasks is not the same as matching a person.
The definitions do the work
OpenAI’s own charter defines AGI as autonomous systems that outperform humans at most economically valuable work. Johnson said that phrasing smuggles in a judgement rather than setting a threshold. “What counts as economically valuable depends on labour markets, institutions, wages, regulation, organisational practice, technological infrastructure, and what societies choose to reward,” she said.
OpenAI has elsewhere described AGI as systems that are generally smarter than humans, which Johnson said creates a different problem. “Which humans? Smarter at what? Across which environments? Under what conditions?” she asked. “Smarter sounds intuitive until you try to turn it into something scientifically measurable.”
The benchmark does not settle it
Brockman leaned on Astra’s score against ARC AGI 3, a test the ARC Prize Foundation uses to compare machine performance with human performance. The foundation ran its own evaluation and was careful about what it would conclude. While it believes Astra represents meaningful progress towards generalisation, it said, it is not claiming the model is AGI.
Walsh said the test “measures a very specific part of AGI” rather than all of human intelligence, and added that no single test exists. Johnson put it more bluntly. A high benchmark score, she said, “can establish strong performance on that benchmark”, but “cannot by itself establish that a system has crossed into a fundamentally different category of intelligence”.
Saturation is worth pausing on. When a benchmark stops separating the best systems from each other, the field usually builds a harder one rather than declaring the problem solved. Walsh made the same point from the other direction: the G in AGI probably means there will never be a single test at all.

Even Altman calls it marketing
The most awkward counterweight to Brockman comes from inside OpenAI. Chief executive Sam Altman recently described AGI as a very poorly defined and irrelevant marketing term, while still suggesting the thing itself is sort of here, or close at least. A company whose president declares an era and whose chief executive calls the label marketing is not sending one message.
That split is not a gotcha, it is a useful signal. When the people selling a system cannot agree on what to call its capability, the label is doing promotional work rather than descriptive work. Brockman himself posted on 3 September that the ARC AGI 3 test was now saturated, which is a claim about a scoreboard, not about a threshold being crossed.
Why Australian businesses should care
This is a procurement problem before it is a philosophy problem. Australian buyers are being asked to sign multi year agreements, move workloads and rebuild processes on the strength of capability claims, and the two people quoted above are among the country’s most credentialled reviewers of exactly those claims. When they say the evidence is not there, that belongs in a vendor assessment.
The practical reading is narrower than the headline. Astra may genuinely be better at driving spreadsheets and forms, and that is worth testing against your own work rather than against a benchmark. Teams weighing one frontier model against another get further from a fortnight of their own tasks than from a leaderboard.
There is a security edge too. A model that controls a browser and fills in forms is acting inside systems that hold customer data, which is a different risk profile from a chat window. That is the same reason enterprise buyers have been pressing vendors on what happens to their data, and it deserves the same scrutiny here.
What to ask a vendor instead
Drop the category question. Whether a model is AGI is not answerable today and would not change a purchase order if it were. The answerable questions are narrower and duller: what tasks did it complete unassisted, on what data, at what error rate, and who checked.
Johnson described the whole dispute as a philosophy of science problem masquerading as a benchmark problem. For a business, the translation is simple enough. Benchmarks tell you a model did well on a benchmark. Only a trial on your own workload tells you whether it does well on yours, and that is the evidence worth collecting while the vendors argue about eras.
