DeepSeek V4 drops with Flash, Pro tiers

DeepSeek-V4-Pro (1.6T MoE) claims parity with GPT-5.4 logic and Claude 4.6 coding tasks at a significantly lower cost, while V4-Flash (284B MoE) targets high-speed agentic workflows. A 90% caching discount ($0.028/1M for Flash) makes it the cheapest 1M-context model on the market. Both models feature a native Thinking Mode and a 384K token maximum output length. The open-weight release maintains DeepSeek's community commitment while providing instant and expert modes for diverse workloads.