DeepSeek V4, Unveiled at Ultra-Low Prices

DeepSeek has unveiled its new V4 model in preview form.

• DeepSeek-V4-Pro supports 1.6 trillion parameters and a 1 million token context.
• The company claims this model is best-in-class among open-source models for agentic coding and reasoning.
• In benchmarks, it emphasized competitiveness against Claude Opus 4.6, GPT-5.4, and Gemini 3.1-Pro.
• The technical report states that V4 is 3 to 6 months behind top frontier models.

Price competitiveness is also a key point.

• V4-Pro output pricing is $3.48 per 1 million tokens.
• V4-Flash is even lower, at $0.28 per 1 million tokens.
• This is significantly cheaper than OpenAI's and Anthropic's rates on the same basis.

On the hardware side, Huawei Ascend chip support stands out.

• DeepSeek collaborated with Huawei to optimize the model to run on Ascend AI processors.
• Huawei stated that Ascend provides "full support" for DeepSeek's models.
• With advanced Nvidia chip access restricted under U.S. export controls on China, China's AI ecosystem shift toward domestic chips could accelerate further.

The article notes that while this release may not shake up the market as dramatically as the previous V3 and R1 did, it once again highlights low-cost, high-efficiency competition and open-source proliferation among Chinese AI companies.