OpenAI’s president, Greg Brockman, recently suggested that its advanced model, known as GPT-6 Astra, could mark the arrival of something called artificial general intelligence, or AGI.
If AI itself is often defined as the ability to simulate human intelligence, OpenAI defines AGI as an AI system that outperforms humans at most economically valuable work – to be better at the majority of digital and cognitive tasks a human can perform. This includes tasks such as writing computer code, organisational activities or managing projects.
For now, at least, it excludes work that involves physical labour, such as agriculture or construction. Though the integration of AI with robotics means such sectors of industry may not be the exclusive domain of humans forever.
This definition by OpenAI focuses on breadth of competency by an AI system. However, other experts maintain that AGI will be characterised by additional cognitive leaps beyond present-day AI models. For example, Demis Hassabis, founder of Google DeepMind, has said that AGI should be able to do “pretty much any cognitive task that humans can do”.
This definition makes AGI more about brilliance and a bit less about depth. Yann LeCun, former chief AI scientist at Meta, believes the term AGI itself is flawed and should be replaced by “human-level AI”.
These represent snapshots from a vigorous ongoing debate over when AGI will arrive and how to recognise it.
VentureBeat magazine reported Brockman acknowledging this at a closed press briefing about GPT-6 Astra in September when he said: “When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognise … it hasn’t played out like that.”
But he added: “For me personally, I do think we’re there.” Just a few days after Brockman made his comments, the chief executive of Anthropic, Dario Amodei, called in an essay titled: “We must pace the frontier” for the AI industry to slow down the pace of its development. Within hours of this essay being published, both Sam Altman and Elon Musk signalled their agreement.
So is Brockman right about AGI being here? The evidence still points in more than one direction and is therefore inconclusive.
The clearest shift with GPT-6 Astra is in how people use it. Earlier models needed detailed instructions at each step. Astra is built for longer jobs where the user sets a goal and some boundaries, and the model works out the sequence itself.
OpenAI’s demonstrations showed the frontier model handling several unrelated tasks at once. The company presents this as moving past the pattern in which a human supplies the next instruction every time.
That is a real change, and it is the part most users will notice first.
One of the benchmarks shows that Astra scored 99.9% on ARC-AGI-3, a test that places an AI agent in unfamiliar interactive games with no instructions and measures how efficiently it works out the rules.
OpenAI estimates that the average human scored 48% when asked to work out the same unfamiliar games. The detail matters here, however. ARC Prize, which runs the benchmark, reported that Astra’s best score through the shared “harness” it gives every model was 62.7%. A harness is a kind of software operating system, or physical environment, within which the AI system works. ARC Prize provides Astra with this software harness.
The 99.9% came from a different harness, built by OpenAI for Astra, which preserves the AI model’s reasoning between requests. Essentially, Astra did better when it was able to reuse previous work and wasn’t “starting from scratch” between requests in terms of its reasoning.
Evaluators disagree
Independent evaluators disagree about how much of an advance GPT-6 Astra represents on its predecessor. The independent research institute Epoch AI places it right at the frontier, alongside other cutting-edge AI models such as Claude Fable 5.1, Kimi K3 and GPT-5.5 Pro. The independent analysis company Artificial Analysis did not rate it as being radically ahead of its predecessor – GPT-5.6 Sol.
Toby Walsh, an AI researcher at the University of New South Wales in Sydney, Australia, described the intelligence in frontier AI models such as GPT-6 Astra and others as “still today very jagged”, telling Al Jazeera that there are simple things that even the best AI models do poorly.
On September 8 2026, OpenAI published a claimed solution to the Navier-Stokes problem, one of seven Millennium Prize problems in mathematics. This has often been reported as Astra’s work. It was not. OpenAI says the proof came from an unreleased internal model, significantly more capable than Astra, running roughly 10,000 AI agents (AI models that perform tasks in response to human instructions) in parallel for about 88 hours.
Astra’s contribution was the formalisation and verification afterwards, which took a further 17 hours. The proof has not yet been checked by the mathematical community.
Amodei’s September essay calling for an AI slowdown remains the strongest recent statement about the risks of frontier models. He wrote that within six to 12 months, a swarm of agents could be capable of taking over the internet with a persistent botnet. This could lead to the catastrophic failure of key systems and would cause damage possibly running into hundreds of billions of dollars.
Claude misuse
It is worth examining Amodei’s warnings. Anthropic’s September threat report sets out examples of how AI is already being used in alarming ways.
According to the Anthropic report, a group in northern Yemen recently used Claude code to build guidance software for rockets and a missile. Russian and Chinese operators automated espionage work. A single French-speaking individual used AI to build a software platform that made personal information about people affiliated with a targeted political movement searchable online. In each case, a human chose the targets and the model supplied the labour.
That is a narrow capability applied at speed and scale under human direction. It is serious, but it is not a description of general intelligence.
Capability and general intelligence are separate things, and they call for different responses. Amodei’s proposed response is to slow frontier development while expanding independent evaluation and coordinating safety standards across countries. A genuinely general system would raise questions that those measures are not designed to answer.
GPT-6 Astra does not settle whether AGI has arrived. Instead, it exposes why that question is becoming harder to answer. Intelligence is increasingly a property of an AI system comprising a model, memory, tools and agents, rather than of a model tested in isolation.
On the narrower question of whether these systems can now cause serious harm, the people running the companies have said plainly that they definitely can.
- is a Lecturer in Business Management and Research, Faculty of Arts, Business, and Social Sciences, University of Wolverhampton
- This article first appeared in The Conversation
