New paper challenges Apple's claim that AI reasoning collapses

A few weeks ago, Apple stirred up the AI world with a paper titled The Illusion of Thinking. Its headline-grabbing conclusion was blunt: even the most capable Large Reasoning Models completely collapse when faced with complex multi-step puzzles.

Now, someone is pushing back hard.

Alex Lawsen, a researcher at Open Philanthropy, just published a rebuttal paper co-authored with Anthropic’s Claude Opus. The title pulls no punches: The Illusion of the Illusion of Thinking. Lawsen argues that Apple did not expose a fatal flaw in AI reasoning, but rather flawed experimental design.

Running out of tokens is not a collapse

Lawsen points out three major design issues in Apple’s original study.

First, token limits were ignored. On Tower of Hanoi tests with eight or more disks, models were actually hitting their output limits. In fact, Claude explicitly stated in some test logs that the pattern continues, but it was stopping to save tokens. Apple’s automated scoring simply flagged this as a reasoning failure.

Second, Apple included mathematically unsolvable puzzles in its River Crossing tests. When models recognized the contradiction and refused to solve them, they were penalized and marked as failed.

Third, the automated evaluation scripts strictly required exhaustive, enumerated move lists. If a model provided a strategic summary or stopped typing, the pipeline graded it as a total collapse.

Code generation tells a very different story

To prove his point, Lawsen ran an alternative test. Instead of forcing models to mechanically type out thousands of moves, he asked them to write a recursive Lua function that prints the solution.

The difference was night and day. Models including Claude, Gemini, and OpenAI’s o3 had no trouble generating correct algorithmic solutions for 15-disk Tower of Hanoi problems. That is far beyond the threshold where Apple claimed they dropped to zero percent accuracy.

As Lawsen put it, the question is not whether reasoning models can think, but whether our benchmarks can tell the difference between reasoning and typing.

Why it matters

If you build with reasoning models, the immediate takeaway is practical: stop asking LLMs to mechanically enumerate huge steps in natural language. If you need complex planning, ask them to write code or algorithmic logic. Writing code sidesteps output token bottlenecks and lets the model demonstrate actual structured logic.

For the AI industry, this debate exposes the fragility of mainstream benchmarks. Apple’s study was widely weaponized by skeptics to claim that post-training reasoning is an architectural dead end. But as Lawsen demonstrated, rigid automated scoring often measures output limits rather than cognitive capability.

We think Lawsen’s rebuttal is sharp and necessary, though it does not completely let frontier models off the hook. Algorithmic generalization remains an open problem. However, before declaring that reasoning models are fundamentally broken, researchers need to ensure their tests are not grading models on their ability to act as high-speed typewriters.