DeepSeek says its new V4 models trail OpenAI and Google by months, not years

China’s DeepSeek is making waves again by claiming its new V4 models are now just months away from catching up to industry leaders like OpenAI and Google.

Roughly a year after its previous release, the Hangzhou-based startup introduced the DeepSeek V4 Pro and V4 Flash preview models on Friday, signaling a massive leap for Chinese AI development.

Performance narrows gap with closed models

The DeepSeek V4 Pro and V4 Flash models are positioned as top-tier contenders. DeepSeek states that V4 Pro leads all open-source models in math and coding benchmarks. Although it lags behind closed systems like Google’s Gemini 3.1 Pro in general knowledge, the performance gap is small. 

DeepSeek estimates that they are now only three to six months behind leading models. The V4 Flash model is designed for speed and efficiency. It offers similar reasoning capabilities to the Pro model but at a lower cost for large-scale use. 

This release follows DeepSeek R1, which some, like Marc Andreessen, considered a turning point in AI. That release showed high-level reasoning could be achieved with less capital, as DeepSeek claimed a training cost of under $6 million. The efficiency of their architecture is apparent, though some analysts are skeptical of that low figure. 

The rapid rise of DeepSeek has led to scrutiny. Because AI has become a central part of the competition between the U.S. and China, these models are under heavy review.

Some regions, including Taiwan, Australia, and parts of the U.S., have restricted the use of earlier DeepSeek models due to data privacy and national security concerns. 

The Stanford AI Index 2026 report confirms that while the U.S. still leads in high-impact patents and model breakthroughs, China has closed the gap in publication volume and industrial applications. 

Competition intensifies across open and closed AI models

DeepSeek continues to use an open-source approach, allowing developers to modify and use their code freely. This puts them in competition with Google’s recently released Gemma 4, which focuses on agent-style workflows and task automation.

As OpenAI refines its closed, enterprise-grade systems, V4 suggests that the choice between open-source accessibility and closed-door performance is becoming more difficult for developers.