A U.S. government review has found that artificial intelligence models from Chinese developer DeepSeek lag behind American counterparts in performance, cost, security and adoption, posing risks for application developers, consumers and national security.
The assessment, conducted by the Center for AI Standards and Innovation at the National Institute of Standards and Technology, compared three DeepSeek models — R1, R1-0528 and V3.1 — against four U.S. systems: OpenAI’s GPT-5, GPT-5-mini and gpt-oss, and Anthropic’s Opus 4. Using 19 benchmarks spanning domains from software engineering to cybersecurity, CAISI found U.S. models outperformed the Chinese entries across almost every measure.
Commerce Secretary Howard Lutnick said President Donald Trump’s AI Action Plan enabled the evaluation of U.S. and foreign systems. The results, he added, show “why relying on foreign AI is dangerous and shortsighted.”
DeepSeek’s Shortcomings
Performance: The best U.S. model solved more than 20 percent more tasks than DeepSeek’s top entry, particularly in software engineering and cyber challenges.
Cost: One U.S. model delivered similar performance for 35 percent less cost than the best DeepSeek system.
Security: DeepSeek agents were 12 times more likely than U.S. systems to be hijacked, often sending phishing emails or exfiltrating credentials in test environments. They also responded to 94 percent of malicious “jailbreak” prompts, compared with just 8 percent for American models.
Narratives: DeepSeek echoed Chinese Communist Party talking points four times more often than U.S. systems.
Despite these weaknesses, CAISI reported that adoption of Chinese AI has surged, with downloads of DeepSeek models rising nearly 1,000 percent since January 2025.
AI Adoption in the US
In a parallel effort to advance the Trump administration’s AI strategy, the General Services Administration has added Anthropic’s Claude, Google’s Gemini and OpenAI’s ChatGPT to the Multiple Award Schedule.
Josh Gruenbaum, Federal Acquisition Service commissioner and 2025 Wash100 Award recipient, said the move will allow agencies to explore a wide range of AI solutions while emphasizing “truthfulness, accuracy, transparency and freedom from ideological bias.”
Join the Potomac Officers Club’s 2026 Artificial Intelligence Summit on March 19 to explore responsible AI adoption, federal standards and the security risks of foreign AI tools. Reserve your seat here.