NIST AI Standards and Innovation Center (CAISI) article
https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro
Part of the content is translated by GPT-5.5 Thinking.
In April 2026, the AI Standards and Innovation Center (CAISI) evaluated DeepSeek V4 Pro (hereinafter "DeepSeek V4"), an open-weight AI model. According to CAISI's evaluation, the capabilities of DeepSeek V4 were estimated to be about 8 months behind frontier models.
DeepSeek V4 is the best performing Chinese AI model that CAISI has evaluated so far. The CAISI evaluation covers areas such as cybersecurity, software engineering, natural sciences, abstract reasoning, and mathematics.
DeepSeek V4 achieved higher scores in DeepSeek's own publicly released evaluations than in the CAISI evaluation. DeepSeek materials state that DeepSeek V4 has capabilities comparable to Opus 4.6 or GPT-5.4, which were released about two months ago. However, according to the CAISI evaluation, which includes non-public benchmarks, DeepSeek V4 performs similarly to GPT-5, which was released about eight months ago.
DeepSeek V4 is more cost-efficient than other models with similar capabilities. Compared to GPT-5.4 mini, the most cost-competitive model in the US, DeepSeek V4 showed higher cost efficiency in 5 out of 7 benchmarks. The cost of DeepSeek V4 in these 7 benchmarks ranged from 53% lower to 41% higher than GPT-5.4 mini.
[Please refer to the link for the rest of the content.]
