U.S. Review Puts DeepSeek V4 Eight Months Behind the Frontier

Overview

• CAISI, a NIST unit, said DeepSeek V4 Pro trails top U.S. systems by about eight months using an exam‑style method across nine tests that include two private datasets.
• Public scores tell a tighter story, with Stanford’s 2026 AI Index reporting a 2.7% gap between leading U.S. and Chinese models on widely used leaderboards.
• After CAISI filtered out pricier or weaker U.S. systems in its cost check, only GPT‑5.4 mini remained and DeepSeek still came out cheaper on five of seven tasks.
• DeepSeek launched V4 with steep introductory prices and claims of near‑frontier results, pitching the model as a low‑cost option for smaller teams.
• Chinese coverage points to likely gains for local chipmakers and fabs, while reports describe tougher oversight of startups and rising competition from Alibaba and ByteDance.