Apple Said AI Can't Think — One Year Later (The Truth)

A year ago, Apple's "Illusion of Thinking" paper went viral with one claim: AI doesn't actually reason. Five days later, Alex Lawsen and Claude Opus 4 showed Apple's tests had a bug. Tower of Hanoi with 15 disks requires 32,767 moves, but the models hit token output ceilings before they could finish. Not a reasoning ceiling. A token ceiling. Twelve months later, in 2026: Claude Opus 4.5, GPT-5, Gemini 3, OpenAI o3 all solve the same puzzles under 5,000 tokens. ARC-AGI-2 leaderboard (May 2026): GPT-5.5 85%, GPT-5.4 Pro 83%, Gemini 3.1 Pro 77%, Claude Opus 4.6 69%. But the asterisk: every score above 50% relies on refinement loops. Search systems using the model as a generator. Not one-pass reasoning. The verdict 12 months later isn't the one most channels gave when they made their hot take in June 2025. It depends what you mean by think. Hypeproof. Hype out. Proof in. Sources: Apple paper: https://machinelearning.apple.com/res... Lawsen + Claude Opus 4 rebuttal: https://arxiv.org/abs/2506.09250 #AI #ReasoningModels #Hypeproof #IllusionOfThinking #ARCAGI #ClaudeOpus #GPT5 #2026AI