Artificial general intelligence: Has AI surpassed human cognitive abilities?
AGI is defined as an AI system that does better than humans at most jobs with high economic value.
OpenAI president Greg Brockman recently suggested that the company's advanced model, called GPT-6 Astra, could mark the arrival of artificial general intelligence, or AGI.
AI is often described as the ability to copy or imitate human intelligence. OpenAI defines AGI as an AI system that does better than humans at most jobs with high economic value. In other words, it would be better than people at most digital and thinking tasks they can do.
These tasks include writing computer code, organising work and managing projects.
For now, this definition does not include jobs that require physical work, such as farming or construction. But as AI becomes more closely linked with robots, these industries may not always be areas where only humans can work.
OpenAI's definition focuses on how well an AI system can handle many different kinds of tasks.
However, other experts believe AGI will need to make further advances in thinking beyond what today's AI models can do. For example, Demis Hassabis, the founder of Google DeepMind, has said that AGI should be able to do “pretty much any cognitive task that humans can do”.
This definition puts more emphasis on being highly intelligent, rather than simply being able to handle many different tasks. Yann LeCun, formerly Meta's chief AI scientist, believes the term AGI is flawed and should be replaced with “human-level AI”.
These views are part of an active debate about when AGI will arrive and how we will know when it has arrived.
VentureBeat magazine reported that Brockman acknowledged this uncertainty during a private press briefing about GPT-6 Astra in September. 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's comments, Anthropic chief executive Dario Amodei called for the AI industry to slow down its development in an essay titled “We must pace the frontier”.
Within hours of the essay being published, Sam Altman and Elon Musk both showed that they agreed.
So is Brockman right that AGI has arrived? The evidence points in different directions, so there is no clear answer yet.
The biggest change with GPT-6 Astra is how people use it. Earlier models needed detailed instructions for every step. Astra is designed to handle longer jobs: users set a goal and some limits, and the model works out the steps needed to reach that goal.
OpenAI's demonstrations showed its advanced model handling several unrelated tasks at the same time. The company says this moves beyond the old approach, where a person had to provide the next instruction at every step.
This is a real change, and it is the difference that most users are likely to notice first.
One benchmark showed that Astra scored 99.9% on ARC-AGI-3. This test puts an AI agent into unfamiliar interactive games without instructions and measures how well and quickly it works out the rules.
OpenAI estimates that the average human scored 48% when trying to work out the same unfamiliar games.
About the author
Waqas Ali is a Lecturer in Business Management and Research in the Faculty of Arts, Business, and Social Sciences at the University of Wolverhampton.
This article is republished from The Conversation under a Creative Commons licence. Read the original article.
However, the details matter. ARC Prize, the organisation that runs the benchmark, reported that Astra's best score using the shared software setup provided to every model was 62.7%.
This software setup is called a “harness”. It is an environment in which an AI system works. ARC Prize provides Astra with this software environment.
The 99.9% score came from a different harness that OpenAI built for Astra. This setup keeps the model's reasoning from one request to the next. In other words, Astra performed better because it could reuse its earlier work instead of having to start its reasoning from scratch between requests.
Evaluators disagree
Independent evaluators disagree about how much better GPT-6 Astra is than the model that came before it.
The independent research institute Epoch AI places Astra among the most advanced AI models, alongside Claude Fable 5.1, Kimi K3 and GPT-5.5 Pro.
The independent analysis company Artificial Analysis did not judge Astra to be far ahead of its predecessor, GPT-5.6 Sol.
Toby Walsh, an AI researcher at the University of New South Wales in Sydney, Australia, said that the intelligence of leading AI models such as GPT-6 Astra remains “still today very jagged”. He told Al Jazeera that even the best AI models struggle with some simple tasks.
On 8 September 2026, OpenAI announced what it said was a solution to the Navier-Stokes problem, one of the seven Millennium Prize problems in mathematics.
This achievement has often been described as Astra's work, but it was not. OpenAI says the proof was produced by an unreleased internal model that was much more powerful than Astra. It used about 10,000 AI agents, which are AI models that carry out tasks in response to human instructions, working at the same time for around 88 hours.
Astra helped afterwards by putting the proof into a formal form and checking it. This took another 17 hours. The mathematics community has not yet checked the proof.
Amodei's September essay calling for AI development to slow down remains the strongest recent warning about the risks of advanced AI models. He wrote that within six to 12 months, a large group of AI agents could be able to take control of the internet through a network of bots that keeps running over time.
This could cause important systems to fail on a huge scale and lead to damage worth possibly hundreds of billions of dollars.
Misuse of Claude
Amodei's warnings are worth examining. Anthropic's September report on threats describes examples of how AI is already being used in worrying ways.
According to the report, a group in northern Yemen recently used Claude Code to build guidance software for rockets and a missile. Russian and Chinese operators used AI to automate spying activities.
A French-speaking individual also used AI to build a software platform that made personal information about people linked to a targeted political movement searchable online. In each case, a human chose the targets, while the AI model did the work.
These examples show a system performing a specific task quickly and on a large scale under human direction. This is serious, but it does not show that the system has general intelligence.
What an AI system can do and whether it has general intelligence are two different questions, and they need different responses. Amodei proposes slowing the development of the most advanced AI systems while increasing independent testing and helping countries agree on shared safety rules.
A truly general AI system would raise questions that these measures are not designed to answer.
GPT-6 Astra does not settle the question of whether AGI has arrived. Instead, it shows why the question is becoming harder to answer.
Intelligence is increasingly a feature of a whole AI system made up of a model, memory, tools and agents, rather than just a model tested on its own.
As for the more specific question of whether these systems can cause serious harm, the people who run the AI companies have clearly said that they can.