DeepSeek releases open-source V4 AI models to rival OpenAI and Anthropic at lower costs

Chinese AI startup DeepSeek has released a preview of its highly anticipated V4 model series, featuring two open-source variants—DeepSeek-V4-Pro and DeepSeek-V4-Flash—that claim to rival top closed-source models from U.S. leaders like Anthropic and OpenAI while slashing costs dramatically. The V4-Pro boasts 1.6 trillion total parameters with 49 billion active, delivering state-of-the-art performance in reasoning, coding, math, STEM, and agentic tasks, according to the company's announcement. Meanwhile, the smaller V4-Flash, with 284 billion total parameters and 13 billion active, offers faster inference and lower pricing at $0.14 per million input tokens and $0.28 per million output tokens.

This launch comes exactly one year after DeepSeek's R1 model stunned the industry in January 2025 by matching proprietary U.S. giants at a fraction of the cost, as reported by Bloomberg. DeepSeek, an offshoot of High-Flyer Capital Management's quantitative analysis firm, positions V4 as a leap forward from its V3.2 predecessor through architectural innovations like novel token-wise compression, DeepSeek Sparse Attention (DSA), and hybrid mechanisms such as Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). These enable a cost-effective 1 million token context length—a default across DeepSeek's services—requiring just 27% of the inference FLOPs and 10% of KV cache memory compared to V3.2 for long contexts.

TechCrunch notes that both models have "closed the gap" with frontier models on reasoning benchmarks, with V4-Pro leading all open-source rivals in world knowledge (trailing only Gemini-3.1-Pro), agentic coding, and competitive tasks—sometimes even outperforming Claude Opus 4.6 on specific agent benchmarks. VentureBeat highlights V4's near state-of-the-art intelligence at 1/6th the cost of models like Opus 4.7 or GPT-5.5, making high-performance AI accessible for developers worldwide. Early tests shared on YouTube and Hacker News praise its real-world prowess in building apps, animations, and agent workflows, with users calling it the best open-source model yet under an MIT license that allows free modification and local deployment.

Not all views are unanimous, however. Bloomberg argues the release fails to narrow America's overall lead in AI, despite DeepSeek's efficiencies and open-source edge. MIT Technology Review emphasizes three key reasons V4 matters: its vastly improved long-prompt handling, open-source availability, and potential to democratize advanced AI tools previously dominated by expensive proprietary systems.

The implications extend beyond benchmarks. Developers can already access V4 via DeepSeek's API, supporting OpenAI and Anthropic formats with thinking/non-thinking modes, and Hugging Face hosts the models for immediate download. This could accelerate adoption in coding, research, and enterprise applications, especially in resource-constrained environments or regions optimizing for chips from Huawei and Cambricon. As competition intensifies ahead of a predicted full 1T-parameter V4 flagship later in 2026, DeepSeek's moves challenge U.S. incumbents and underscore China's growing role in the open-source AI revolution, affecting global innovators, businesses, and policymakers alike.