User tests indicate that OpenAI’s GPT-4 model exhibits abnormal performance under specific conditions, such as responding without thinking, fast generation speeds, and poor instruction following. The author resolved the issue by switching proxy nodes, speculating that OpenAI routes models to the lower-performance gpt-4o-mini based on Juice values. Data from the ChatGPT K12 Teacher Edition Dashboard reveals potential mechanisms for AI model performance control. The author notes that degradation occurs in three levels, with severe cases uniformly routed to gpt-4o-mini. This discovery has important implications for AI users and developers, reminding them to pay attention to model performance fluctuations and resource management strategies.
