Apple says its most powerful AI models still can’t think through hard problems

Apple just dropped a surprising truth bomb about AI’s so-called “reasoning” powers. In a new research paper published last week on Saturday, Apple scientists say even the smartest AI models, especially those designed to think, start to unravel when faced with problems that are just too complex. Instead of trying harder, they give up.

That is not enough

The paper, titled “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity,” explores how Large Reasoning Models (LRMs) and regular Large Language Models (LLMs) behave when challenged with tasks of increasing complexity. Spoiler alert, they do fine until things get really tricky.

ALSO READ: iOS 26 is bringing a bold new “Liquid Glass” look, just ahead of iPhone’s biggest redesign yet 

To test the models, Apple researchers turned to brain teasers, literally. They used puzzles like the Tower of Hanoi, a problem often used to test logic in kids. It starts simple, moving three disks across pegs. But things escalate fast when you add more disks, and that’s where the AI began to wobble.

In the experiment, researchers tested both standard LLMs (Claude 3.7 Sonnet and DeepSeek-V3) and their “thinking” versions (Claude 3.7 Sonnet with Thinking and DeepSeek-R1), each allowed to use up to 64,000 tokens, which is a massive compute budget. They were tested across three task types: low (3 disks), medium (4-10), and high complexity (11-20).

The results showed that both model types performed similarly at lower levels. With medium difficulty, LRMs used the extra compute to do better. But once the task crossed into high complexity, they both broke down. The logic vanished, shortcuts were taken, and some models simply gave up.

This wasn’t a one-off. The paper says similar patterns emerged across other classic puzzles like River Crossing, Blocks World, and Checkers Jumping.

All in all, Apple’s findings echo what many in AI already suspect. It is that these models might look smart, but when pushed to truly “think,” they still hit a wall.