Chinese AI lab DeepSeek released V4, its first major flagship model since the breakthrough R1 release that transformed the company into China's most recognizable AI player. The new model comes in two variants: V4-Pro for complex coding and agent tasks, and V4-Flash as a lighter, faster alternative. Both versions handle 1 million tokens of context—enough to fit the entire Lord of the Rings trilogy plus The Hobbit.

The pricing is aggressively competitive. V4-Pro charges $1.74 per million input tokens, while V4-Flash runs just $0.14—fractions of what OpenAI and Anthropic demand. DeepSeek claims V4-Pro matches Anthropic's Claude-Opus-4.6, OpenAI's GPT-5.4, and Google's Gemini-3.1 on major benchmarks, outperforming open-source rivals like Alibaba's Qwen-3.5 on coding, math, and STEM problems.

V4 introduces architectural innovations that sharply improve memory efficiency. The model uses selective attention, compressing older information while focusing on relevant context, cutting computing costs to just 27% of the previous V3.2 for million-token contexts. Perhaps most notably, V4 is DeepSeek's first model optimized for Huawei's Ascend chips rather than Nvidia hardware—a strategic pivot reflecting US export restrictions and China's push for semiconductor self-reliance.