DeepSeek V4 nears GPT-5-level performance per US CAISI evaluation

DeepSeek V4 (1.6T params, MIT license) was evaluated by the U.S. AI Standards and Innovation Center as performing similarly to OpenAI's GPT-5 (Aug 2025), making it the best Chinese AI model — though still trailing the newest U.S. models by ~8 months. Cost efficiency was a noted strength.