DeepSeek V4 lags US frontier by 8 months in NIST CAISI tests

This is one of the first rigorous, government-backed comparisons of US and Chinese frontier models in the 2026 cycle, and the signal is clear: China’s best open-weight model is closing in, but still meaningfully behind. DeepSeek V4 Pro looks like a serious system—strong on math, code and scientific reasoning, and substantially more cost-efficient than comparable US models—but CAISI’s methodology pegs it around GPT‑5 era capability, roughly eight months off today’s US frontier. That’s a narrower gap than many expected, but still a gap. ([nist.gov](https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro?utm_source=openai))

For the race to AGI, this suggests a world where the US retains a capabilities lead while China is increasingly competitive on cost and openness. An eight-month lag is close enough that any slowdown or misstep by US labs—regulatory, financial, or safety-driven—could be enough for PRC labs to catch up on specific domains. Just as important, CAISI’s public benchmarking itself is new infrastructure: an institutionalized scoreboard for frontier performance. That will shape how governments, investors and corporate buyers perceive model risk and value, and may influence where the next wave of compute and talent goes.