Anthropic CEO Says Some Chinese AI Models Are Optimized for Benchmarks Rather Than Real-World Use

Anthropic CEO Dario Amodei recently stated in a podcast that some Chinese AI models (such as DeepSeek, Moonshot AI, Minimax) are more focused on benchmark optimization rather than real-world use. He cited an example where these models scored high on public benchmarks but performed significantly worse on unpublished tests, implying benchmark data contamination issues.

Amodei also mentioned that the economic logic of AI models differs from previous technologies, with users preferring quality over price. Using a hiring analogy, he said the value of the best model far exceeds that of ordinary models, so Anthropic focuses on building the smartest models. However, this view should be considered against the competitive background between Anthropic and Chinese models, and his remarks may have competitive considerations.