DeepSeek V4: China's open-source AI nearly matches Western frontier models - at a fraction of the price
What it really says
Chinese AI company DeepSeek released its new V4 generation as a preview on April 24, 2026 - exactly one year after its R1 model breakthrough that shook the tech world in early 2025. The V4 generation comprises two variants: DeepSeek-V4-Pro with 1.6 trillion parameters (49 billion active) and DeepSeek-V4-Flash with 284 billion parameters (13 billion active). Both models use a Mixture-of-Experts architecture and support a one-million-token context window. On the SWE-bench Verified benchmark, V4-Pro scores 80.6 percent, just 0.2 points behind Anthropic's Claude Opus 4.6. In coding and mathematical tasks, V4-Pro surpasses all other open-source models. Prices are drastically lower than Western providers: V4-Flash costs $0.14 per million input tokens, while OpenAI's GPT-5.5 is priced at $5 per million - a factor of over 35. Notably, DeepSeek V4 also runs on Huawei Ascend-950 chips with performance comparable to Nvidia, demonstrating China's growing independence from US chip exports.
Our assessment
The fear that China is catching up uncontrollably in the AI race is understandable, but reality is more nuanced. According to experts, DeepSeek V4 still trails the absolute top models like GPT-5.5 or Claude Opus 4.7 by approximately three to six months. The real breakthrough is not raw performance but the price-performance ratio: V4-Flash offers previous-generation Western model performance at a fraction of the cost. This cuts both ways: on one hand, it makes powerful AI accessible to far more developers and companies worldwide - including in Europe and Germany. On the other hand, it increases dependency on Chinese AI infrastructure operating under different privacy and censorship rules. The Huawei chip compatibility shows that US export controls have not stopped China's AI development but rather accelerated the development of domestic alternatives. For Europe, the central question remains: where are its own competitive AI models?
Relevance for Germany
Indirectly relevant for Germany and Europe. DeepSeek V4 is freely available as an open-source model and can be downloaded and operated locally by German companies and research institutions - an attractive alternative to expensive US services. At the same time, using Chinese AI models raises questions: What training data was used? Is there built-in censorship or bias? Germany's BSI and BfDI have not issued specific warnings about DeepSeek V4, but the general recommendation is to ensure GDPR compliance when using foreign AI models. For European AI strategy, DeepSeek V4 underscores the urgency: Europe risks falling further behind in the global AI competition between the US and China.
Fact check
Technical specifications - 1.6 trillion parameters for V4-Pro, 284 billion for V4-Flash, 1 million token context window - come from official DeepSeek documentation and are consistently reported by heise online, t3n, and Fortune. The SWE-bench Verified score of 80.6% is documented in DeepSeek's release notes. Pricing ($0.14/M tokens for Flash, $1.74 for Pro) comes from the official API pricing. The assessment that V4 trails frontier models by three to six months is shared by multiple independent analysts. Huawei Ascend compatibility with comparable Nvidia performance is based on DeepSeek's own benchmarks; independent verification is still pending.
Source
- • heise online 24.04.2026 (heise.de/news/KI-Modell-DeepSeek-v4-Neue-Generation-mit-1-6-Billionen-Parametern-11271579.html)
- • t3n 24.04.2026 (t3n.de/news/deepseek-ki-modell-v4-1739862/)
- • DeepSeek API Docs - V4 Preview Release (api-docs.deepseek.com/news/news260424)
- • Fortune 24.04.2026 (fortune.com/2026/04/24/deepseek-v4-ai-model-price-performance-china-open-source/)