"Car Wash" test with 53 models
opper.ai
February 23, 2026
9 min read
🔥🔥🔥🔥🔥
64/100
Summary
The car wash test evaluates AI reasoning by asking whether to walk or drive 50 meters to a car wash. Most leading AI models, including Claude Sonnet 4.5, GPT-5.1, Llama, and Mistral, fail to provide the correct answer, which is to drive.
Key Takeaways
- Only 11 out of 53 AI models correctly answered the car wash test, with 42 models incorrectly suggesting to walk.
- The models that consistently passed the test across 10 runs were Claude Opus 4.6, Gemini 2.0 Flash Lite, Gemini 3 Flash, Gemini 3 Pro, and Grok-4.
- OpenAI's GPT-5 failed 30% of the time in the consistency test, providing incorrect reasoning related to fuel efficiency instead of the need for the car at the car wash.
- 33 models, including all Llama and Mistral models, failed to answer the car wash test correctly in all attempts.
Coverage & discussions
Hacker News · opper.ai
Community discussions
Community Sentiment
Positives
- The 'Car Wash Test' highlights significant gaps in AI reasoning, prompting discussions on how models interpret context and ambiguity in questions.
- Models like Sonnet 4.6 demonstrate high common sense reasoning capabilities, suggesting potential for improved performance in understanding nuanced queries.
Concerns
- Many AI models struggle with basic reasoning tasks, indicating a fundamental flaw in their design and prompting concerns about their reliability in real-world applications.
- The human baseline for evaluating AI responses is flawed, as it lacks proper screening and does not require reasoning, potentially skewing results.
