Astra Meets Ad Astra: Has the AGI Era Begun?

OpenAI’s new GPT-6 Astra can use computers, conduct research, write code and carry out complex tasks with considerably less human intervention. Greg Brockman believes it may be reasonable to speak of the beginning of an AGI era. I’m less interested in declaring that moment than in what Astra tells us about the way our relationship with AI is changing.

I have to admit that the name caught my attention first.

When you publish a journal called Ad Astra and OpenAI releases a model called Astra, you are going to notice. Astra is Latin for “stars”; ad astra means “to the stars”. OpenAI has not explained publicly why it chose the name, so there is no great revelation to offer here. It was simply a coincidence that made me curious enough to read further.

What I found was considerably more interesting than the name.

GPT-6 Astra is OpenAI’s latest model, and one of the important differences is its ability to work directly with computers. It can navigate websites and software, conduct research, fill in forms, work with documents and spreadsheets, write and run code and carry out tasks involving several steps.

We have heard lists like this before, of course. Every new generation of AI arrives with a collection of benchmarks and demonstrations showing that it is faster, more capable or better at reasoning than the last. It is easy to become numb to them.

But computer use is worth paying attention to because it changes something quite practical.

Until recently, most people experienced generative AI as an exchange. We asked for something and the model produced it. Even when the answer was very good, there was still a person sitting between the AI and the outside world. Someone had to send the email, enter the information, open the spreadsheet, run the programme or decide what happened next.

That gap is getting smaller.

Astra can increasingly carry out some of those actions itself. This doesn’t mean that it can suddenly replace every person working at a computer, and some of the more enthusiastic commentary around AI would benefit from remembering that. But it does mean that the conversation is moving beyond what a model can tell us towards what we are willing to allow it to do.

Greg Brockman, OpenAI’s co-founder and president, has gone further. Speaking around the launch of Astra, he suggested that it was “not unreasonable” to think that we are entering the AGI era.

I find the wording interesting because AGI — artificial general intelligence — remains a surprisingly slippery concept for something discussed with such certainty. There is no single definition accepted by researchers, companies and governments. Depending on whom you ask, it might mean an AI capable of performing most intellectual tasks at human level, an AI that can adapt successfully to unfamiliar problems, or a system capable of doing a substantial proportion of economically useful work.

For years, AGI was imagined as a fairly obvious event. At some point a threshold would be crossed and we would know that we had entered a different technological age.

I am beginning to wonder whether we will recognise it quite so neatly.

It seems equally possible that there will be no single AGI moment. Models will become better at reasoning, then better at using tools, then better at navigating computers, then more reliable at carrying out longer tasks. Each development will be announced separately and debated for a few days. Eventually we may look around and realise that the definition of an AI assistant has changed rather substantially while we were arguing about terminology.

Astra is certainly impressive by the measures being used to test it. Greg Kamradt of the ARC Prize Foundation said that it surpassed their human action-efficiency baseline on 96 per cent of ARC-AGI-3 levels, effectively reaching human parity on that benchmark. Greg Burnham of Epoch AI described the release rather neatly: “The story is: end of one era, start of another.”

I would keep a little distance from both the excitement and the numbers.

Benchmarks tell us something useful about capability, but professional life is not a benchmark. People work with bad instructions, incomplete information, office politics, conflicting priorities and consequences that cannot always be reduced to a score. Judgement remains difficult to measure, and confidence is not the same thing as competence — for humans or machines.

There is another part of the Astra release that received less of the glamorous attention but interested me more. Astra is the first OpenAI model to reach the company’s “Critical” capability threshold for cybersecurity.

According to OpenAI, with the necessary tools and access, Astra is capable of identifying previously unknown vulnerabilities and developing ways to exploit highly protected systems without needing detailed human guidance at every stage. OpenAI has consequently placed additional safeguards and restrictions around the model’s most advanced cybersecurity capabilities.

This is where the questions around autonomy stop being abstract.

The very thing that makes an AI agent useful — giving it enough freedom to complete a task without asking us what to do every thirty seconds — also creates the problem of how much freedom it should have.

There is a considerable difference between trusting a model to summarise a document and trusting it to take actions inside a system. There is another considerable difference between allowing it to book an appointment and allowing it to execute code or interact with sensitive infrastructure. We will presumably become comfortable with some of these things very quickly and remain deeply uncomfortable with others.

Where that boundary settles will matter more to most people than whether we eventually agree to call Astra AGI.

The same is true of the debate about work.

Every significant AI release now produces predictions about which professions are about to disappear. I have never found that framing particularly convincing. A profession is rarely just a collection of tasks that can be listed on a spreadsheet. There are relationships, experience, judgement and responsibility involved, often in ways that become visible only when something goes wrong.

But it would be equally naïve to pretend that nothing substantial is changing.

If an AI can conduct the first round of research, prepare the spreadsheet, assemble a presentation, test the software and move information between the programmes people use every day, it does not have to replace an entire profession to alter how that profession works.

Some tasks will disappear. Others will become faster. New ones will appear. And people may find that their value lies less in producing the first version of something and more in knowing whether that version is any good.

That last point interests me particularly.

As AI becomes better at execution, judgement becomes more valuable, not less. Someone still has to decide what is worth doing, which information deserves to be trusted, when an efficient solution is the wrong solution and who is responsible when a decision has consequences.

Those are not engineering questions. They are human ones.

And perhaps that is why I am reluctant to become too absorbed in whether Astra qualifies as AGI. The terminology is fascinating, but it can distract us from decisions that are already arriving.

Companies will have to decide what they are prepared to automate. Governments will have to decide where intervention is necessary. Professionals will have to decide what they are comfortable delegating. Individuals will make smaller versions of the same decision every time an AI asks for access to another application, account or part of their lives.

None of this means we should approach Astra with fear. I don’t. There is something genuinely exciting about seeing machines become capable of things that would have sounded implausible only a few years ago. Curiosity about technology is one of the reasons I started Ad Astra Journal in the first place.

But curiosity does not require credulity.

We can be interested in what a technology makes possible while still asking who controls it, who benefits from it and where responsibility sits when it acts on our behalf.

Which brings me back, finally, to that name.

Astra. Stars.

Ad Astra. To the stars.

I don’t know whether OpenAI intended any grand symbolism when it chose it, and until the company says otherwise, there is little point pretending that we do.

But I like the coincidence.

The language of the stars has always carried something optimistic with it: curiosity, distance, exploration, the possibility that there is still more to discover. Those are ideas I believe in.

The important thing is not to confuse exploration with inevitability. Being able to go somewhere has never meant that we should stop thinking about how we get there.

Has the AGI era begun? Perhaps. Ask me again when we have had more time to live with Astra outside demonstrations and benchmarks.

For now, I find another question more useful.

Not simply what can AI do now?

But what do we want it to do for us — and what would we rather continue doing ourselves?

That is the conversation I suspect will last much longer than the launch.