DeepSeek V4: The Open-Source Model Frontier Labs Feared (helloai.com)

DeepSeek has launched DeepSeek V4, an open-source AI model that dramatically disrupts pricing standards in the AI/ML community. Priced at just $0.30 per million output tokens, it offers an 83-to-100x cost advantage over proprietary models like Claude Opus 4.7 and GPT-5.5, while achieving an impressive 80.6% score on SWE-bench Verified benchmarks. As it ships under an MIT license on Hugging Face, this model not only represents a more accessible entry point for developers but also signals a significant shift in the competitive landscape for AI coding applications. The architecture of DeepSeek V4-Pro is notable for its efficiency; it utilizes a 1.6-trillion-parameter mixture-of-experts (MoE) model that activates only 49 billion parameters per token, significantly lowering inference costs and making the model feasible for teams operating their own GPU infrastructure. The model’s performance, including a LiveCodeBench Pass@1 score of 93.5, positions it competitively against existing closed models, potentially reshaping procurement strategies and expectations around coding intelligence. However, the concerns around benchmark transparency and the potential implications of sourcing from a Chinese lab persist, encouraging teams to weigh the trade-offs between cost savings and data governance carefully.