DeepSeek V4-Pro hits 150+ tok/sec/user on GB200 NVL72, rivals frontier at fraction of cost

DeepSeek V4-Pro, released via API on April 24, packs 1.6 trillion total parameters and a 1M-token context window, and delivers 150+ tokens/sec/user on NVIDIA's GB200 NVL72 platform. It reportedly rivals Claude Opus 4.7 and GPT-5.5 in coding-heavy agentic workflows at a small fraction of the cost, and V4 was reportedly trained for $7.8M.