Have we entered the AGI era?

It started with three words

On 3 September 2026, OpenAI released GPT-6 Astra. At the end of the launch briefing, OpenAI president Greg Brockman said:

It was an extraordinary phrase from the president of a company that has spent years making AGI central to its mission. But the detail matters.

Brockman was not announcing that a universally agreed technical threshold had been crossed. He described AGI as a fuzzy concept and left room for people to decide whether Astra itself qualifies.

His broader point was more interesting. He suggested that if people look back in a few years and ask when AGI was born, they may identify this period, and possibly this model, as the point at which something changed.

That is a different claim from saying AGI has arrived. And it may ultimately be a more useful one.

An era doesn’t need a starting gun

We tend to talk about AGI as though someone will eventually announce it. One day we have AI. The next day we have AGI. But there is no accepted test that would make that happen.

Some definitions focus on whether AI can outperform humans across economically valuable work. Others require something closer to the breadth and flexibility of human cognition. Some emphasise the ability to learn new things. Others care more about autonomy, reasoning or the ability to operate reliably in unfamiliar situations.

A system could therefore satisfy one person’s definition of AGI while falling well short of another’s.

Seven definitions in current use, and what each would require →

That makes the idea of an AGI era interesting. An era can begin before everyone agrees on the date.

Something has changed

Whatever label we eventually put on today’s systems, the argument is no longer about whether AI can produce convincing text or generate a picture.

Frontier systems can now work across text, images, software, research, mathematics and other professional tasks. They can use tools, operate computers, carry out increasingly long sequences of actions and contribute to scientific and technical work.

They can also still make surprisingly basic mistakes. That combination matters.

The systems are becoming broader without becoming uniformly reliable. They can appear remarkably general in one situation and strangely limited in another.

That is one reason the AGI debate has become so difficult. The old distinction between narrow AI and general intelligence is becoming less useful at exactly the moment when the remaining differences matter most.

Why the narrow-versus-general distinction is breaking down →

And now DeepMind is building an institution around what comes next

Two weeks after Brockman’s “AGI era” remark, another significant development arrived. Google DeepMind launched the DeepMind Institute, led by Demis Hassabis, Shane Legg and James Manyika.[TNW]

Its purpose is not to announce that AGI has been achieved. It is to examine what increasingly capable AI, and eventually AGI, could mean for society. The Institute’s opening work looks at questions including AI safety, deception, economic disruption, scientific progress and how society might prepare for increasingly general systems.[The Decoder]

The language around AGI is changing

For years, AGI was largely discussed as a destination. When will we reach AGI? What would AGI be able to do? Which company will build it first?

Increasingly, the discussion is shifting towards another set of questions. What happens as systems become more general? How do we know which capabilities matter? How should increasingly autonomous systems be evaluated? How do we prepare businesses, governments and workers? What risks become important before full AGI exists? And at what point does arguing about the label become less important than responding to what the systems can actually do?

That doesn’t mean everyone suddenly agrees that AGI is close. They don’t. But the conversation is moving from a hypothetical future towards an increasingly practical present.

AI is also beginning to help develop AI

There is another reason this period could eventually look significant. AI is increasingly being used inside AI research itself.

Models can help researchers write and analyse code, generate experiments, evaluate outputs, search technical literature and work through difficult mathematical and scientific problems.

That is not the same thing as an AI system independently redesigning itself. The distinction is important. But once AI contributes meaningfully to the research process that produces better AI, the relationship between one generation and the next starts to change.

The people building the technology are no longer working without it. AI has become part of the machinery of AI development. If that contribution becomes substantially more capable, the speed of progress could change with it.

But the systems are still unreliable

This is the part that gets lost when AGI is discussed as a race. Capability and reliability are not the same thing.

A system might solve an extremely difficult mathematical problem and then fail at a much simpler task. It might complete a long piece of work successfully sometimes, but fail often enough that a human still needs to supervise it. It might produce something genuinely useful while remaining poor at recognising when it is wrong. It may also perform extremely well on a benchmark without demonstrating the flexibility that people usually associate with general intelligence.

These are not minor details. They are some of the biggest unresolved parts of the AGI question.

The six ways a benchmark score and a headline come apart →

The missing capabilities matter more than the label

If we want to know whether the transition towards AGI is genuinely happening, there are several things worth watching.

Progress on those dimensions would tell us far more than whether somebody has attached the letters AGI to a product launch.

So, are we in the AGI era?

There isn’t a measurement that can settle it.

If by AGI era we mean a period in which universally accepted human-level general intelligence already exists, the evidence does not establish that.

If instead we mean a period in which increasingly general AI systems are beginning to alter work, research and the development of AI itself, while the organisations building them start preparing explicitly for what comes next, the case is much stronger.

That is why the word era may turn out to be more useful than arrival.

The transition may not happen in one release. We may get systems that become more capable across more domains while retaining strange weaknesses. They may become economically transformative before satisfying everyone’s cognitive definition of AGI. Some capabilities may arrive years before others.

There may never be a universally accepted morning when the world wakes up and agrees: yesterday was AI, today is AGI.

We may only know afterwards

Technological eras are often easier to identify in retrospect. The significance of a particular model, breakthrough or year becomes clearer once we can see what followed it. That may happen with AGI too.

Perhaps GPT-6 Astra will eventually look like an important step but not the beginning of anything fundamentally new. Perhaps another system will provide a much clearer dividing line. Or perhaps the entire search for one decisive AGI moment will eventually look misguided.

We may discover that general intelligence arrived unevenly: one capability at a time, across several generations of systems, until the cumulative change became impossible to ignore.

That is why Brockman’s three words matter without having to accept them as a declaration that AGI has been achieved. And it is why DeepMind creating an institution around preparing for AGI matters too.

Neither event settles the argument. Together, they show how much the argument itself has changed.

Who is saying AGI has arrived, updated as positions change →

How would we know if AGI had arrived? →

Who is winning the AGI race? →

What is AGI? The plain-English version →