A Five-Year-Old AGI Test Just Beat AI and Exposed a Hard Truth

For an industry racing toward artificial general intelligence, a new benchmark just delivered a surprisingly blunt reality check.

Some of the most advanced AI models in the world systems from companies like OpenAI, Google, and Anthropic just failed a test that humans, including young children, can solve with ease. And not by a small margin.

On the ARC-AGI benchmark, designed to measure real reasoning rather than memorisation, leading models scored below 1%, while humans were able to solve nearly all of the tasks. That’s not a performance gap, it’s a structural one.

What makes this result so uncomfortable for the AI narrative is the nature of the test itself. Unlike most benchmarks that reward pattern recognition over massive datasets, ARC-AGI focuses on something much harder which is the ability to understand a completely new problem, infer its rules, and solve it with minimal context. It’s the kind of thing humans do instinctively, often without even realising it.

In simple terms, this is what researchers mean by artificial general intelligence (AGI), a system that can take on unfamiliar problems, figure them out from first principles, and adapt in real time, without needing to be retrained for every new task. It’s the difference between an AI that knows a lot and one that can truly think.

AI, for now, does not.

That distinction cuts to the core of what today’s systems actually are. Despite impressive gains in writing, coding, and automation, most large models still depend heavily on learned patterns. When those patterns don’t apply when the problem is unfamiliar or abstract, they tend to break down quickly.

This is where the hype around AGI starts to look fragile. Over the past year, the narrative has been one of relentless progress, with claims that AI is approaching or even achieving general intelligence. But results like this suggest something more nuanced: AI is getting better, but not necessarily more general.

And that difference matters more than it sounds.

A system that can summarize documents or generate code is useful. A system that can walk into a completely new situation, figure it out, and adapt in real time is something else entirely. That’s the threshold for true general intelligence, and it’s also the point where AI begins to replace more complex human roles.

Right now, that threshold still looks distant.

The timing makes the contrast even sharper. This benchmark lands at a moment when AI investment is surging, valuations are climbing, and expectations are being set at historic highs. Against that backdrop, a test that even a child can outperform doesn’t just feel ironic, it feels grounding.

It suggests that the hardest part of AI isn’t scale, or data, or compute. It’s generalisation—the ability to learn something new without prior training.

And until that problem is solved, the gap between today’s AI and true AGI isn’t just about performance.

It’s about understanding.