CarWashBench flunks most frontier models

Tiny benchmarks can be noisy, but this one is a sharp gut-check for whether flagship "reasoning" models are actually reasoning or just pattern-matching.

• –The benchmark is extremely small at just two questions, so the leaderboard is more signal flare than final verdict.
• –Even so, near-total failure from multiple top-tier models is notable because the task targets everyday common-sense reasoning rather than specialized knowledge.
• –Gemini 3.1 Pro standing well above the field suggests some models are better at escaping superficial heuristics on deceptively simple prompts.
• –If the author expands the question set without losing the adversarial framing, this could become a useful lightweight reasoning stress test.