Connect with us
Image by Google Gemini, based on a prompt by Gadget.

Artificial Intelligence

AI is now superior to
humans – or is it?

OpenAI’s announcement of the Astra model on Thursday came with a breathtaking claim of achieving AGI, but ironies abounded, writes ARTHUR GOLDSTUCK.

It should have been a watershed moment in the history of artificial intelligence. When OpenAI president Greg Brockman ended a briefing with reporters on Thursday by saying, “Welcome to the AGI era”, he was implying that the company’s new Astra model had achieved the holy grail of human-level intelligence.

Artificial General Intelligence (AGI) is the idea of an AI system that can perform most intellectual tasks at or above human level across many fields, instead of being exceptionally good at a narrow set of jobs.

It represents such a powerful milestone, that it is utterly unconvincing to hear the claim from the mouth of a human being. If OpenAI really believes this is AGI, it could have proven this by allowing GPT-6 Astra conduct the briefing itself.

The company calls it its most intelligent and aligned model and says it can operate computers, browse the web and carry out demanding professional work. It can fill in forms, work through spreadsheets, create software and conduct online research. Sp why not put Astra on a screen in front of the reporters, give it the launch documents and safety material, and let it take questions about its own capabilities?

That would have been a far stronger demonstration than another set of benchmark charts. The measures OpenAI has published are impressive enough on their own. On a test called Agents’ Last Exam, which tests complex professional tasks in real software, it scored 59.3%: up from 53.6% for the current flagship model, GPT-5.6 Sol. On AutomationBench, its 41.4% score was more than double Sol’s 18.1%. OpenAI also says it can complete computer tasks substantially faster than its predecessor.

Admittedly, Astra is the first OpenAI model to reach the Critical level under the company’s Preparedness Framework. OpenAI says that, with the right tools and access, it can find previously unknown security flaws and develop ways to exploit well-protected systems without a person guiding every step. It is one reason the company has restricted access to some of its strongest cyber functions.

OpenAI’s Charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”. That definition almost screams for a demo in the real world.

OpenAI’s own launch page is more cautious. It calls Astra the company’s most intelligent and aligned model, and describes advances across computer use, software engineering, science and professional work. It does not declare that Astra is AGI, but the implication appears in the name of the ARC-AGI benchmark.

OpenAI gives Astra a spectacular 99.9% score on ARC-AGI-3, a benchmark designed to test how well a system learns to deal with unfamiliar environments. ARC Prize, which runs the benchmark, reports 62.7% for Astra using its standard testing and 99.9% with an OpenAI-specific adapter that preserves the model’s internal reasoning state between requests. ARC Prize calls the performance a major milestone but explicitly says that saturating (scoring close to the maximum possible score) its benchmark would not prove AGI.

Artificial Analysis gives Astra a score of 61 on its Intelligence Index, level with GPT-5.6 Sol and behind Anthropic’s Claude Fable 5.1 at 66. Astra makes much stronger gains in coding-agent work and in using computers. But he new model appears to be advancing fastest in the area most relevant to replacing chunks of human work: actually doing things on a computer.

Thar mean the ordinary lay uiser may well experience it as AGI, but that is not the same thing as actual AGI.

Brockman himself softened the declaration when reporters pushed him on it. He described AGI as a “mission concept or spiritual concept”, left it to readers to decide if Astra qualified, and said that, personally, he thought it did. That is a rather loose description for something OpenAI has always defined in economic terms.

Either way, There are many reasons Astra should have conducted the press briefing.

Mainly, it would have tested a different kind of ability, since reporters do not follow benchmark scripts, they challenge assumptions, ask for evidence and change direction when an answer raises a new question. Astra would have to know the launch material, keep its answers consistent with company policy and recognise where it lacked enough information to answer.

Perhaps it is because senior executives earn a large part of their salaries for precisely that kind of judgement, and no one at OpenAI wants to declare their job redundant, even as they do so for countless others in management.

OpenAI’s own description of Astra gives a clue to why Brockman was still the one making the announcement. Its launch material says that, when a task reaches a consequential decision, the model “waits for your input”.

In other words, move along human, nothing to see here.

 Arthur Goldstuck is CEO of World Wide Worx, editor-in-chief of Gadget.co.za and GadgetWings, and author of “The Hitchhiker’s Guide to AI – The African Edge”.

Subscribe to our free newsletter
To Top