DeepSeek V4’s real breakthrough is cost-efficient long-context intelligence: it makes million-token reasoning cheaper and pushes open models closer to frontier systems.