US AI models outperform Chinese rival DeepSeek
CAISI found Chinese AI models from DeepSeek lag behind US models in performance, cost, security, and adoption, raising concerns for developers and national security.
The National Institute of Standards and Technology’s Centre for AI Standards and Innovation (CAISI) found AI models from Chinese developer DeepSeek trail US models in performance, cost, security, and adoption.
Evaluations covered three DeepSeek and four leading US models, including OpenAI’s GPT-5 series and Anthropic’s Opus 4, across 19 benchmarks.
US AI models outperformed DeepSeek across nearly all benchmarks, with the most significant gaps in software engineering and cybersecurity tasks. CAISI found DeepSeek models costlier and far more vulnerable to hijacking and jailbreaking, posing risks to developers, consumers, and national security.
DeepSeek models were observed to echo inaccurate Chinese Communist Party narratives four times more often than US reference models. Despite weaknesses, DeepSeek model adoption has surged, with downloads rising nearly 1,000% since January 2025.
CAISI is a key contact for industry collaboration on AI standards and security. The evaluation aligns with the US government’s AI Action Plan, which aims to assess the capabilities and risks of foreign AI while securing American leadership in the field.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
