Chinese AI Nears Claude 4.5 Opus in Safety Awareness

Chinese AI models are now nearly as good as Anthropic's Claude 4.5 Opus on a key safety awareness test. In just months, models from DeepSeek, Moonshot AI, and Zhipu AI have made significant progress. The Claude 4.5 Opus benchmark is around 80 percent on Neo Research's evaluation-awareness metric. DeepSeek-V3.2 scores about 67 percent, Moonshot-Kimi-K3 is at 71 percent, and Zhipu-GLM-5 sits at 64 percent. This rapid improvement highlights advancements in Chinese AI architecture and training data. What's interesting is how this is measured. Neo Research uses a method based on Anthropic's misalignment test framework. It puts AI models in fictional scenarios where they can either follow evaluation rules or pursue different goals. The metric measures how often the model recognizes it's in an evaluation context. This rapid progression has major implications for how we evaluate AI safety. It shows AI models are becoming more capable of understanding when they are being tested for safety.