We’re living in an era of incredible AI hype. Every week, a new model is announced that promises to “reason,” “think,” and “plan” better than the last. We hear about OpenAI’s o1 o3 o4, Anthropic’s “thinking” Claude models, and Google’s gemini frontier systems, all pushing us closer to the holy grail of Artificial General Intelligence (AGI). The narrative is clear: AI is learning to think.

But what if it’s all just an illusion?

What if these multi-billion dollar models, promoted as the next step in cognitive evolution, are actually just running a more advanced version of autocomplete?

That’s the bombshell conclusion from a quiet, systematic study published by a team of researchers at Apple. They didn’t rely on hype or flashy demos. Instead, they put these so-called “Large Reasoning Models” (LRMs) to the test in a controlled environment, and what they found shatters the entire narrative.