DeepSeek released V4-Pro (1.6T parameters, 1M context) on April 24, 2026, pricing output at $3.48/M tokens versus OpenAI's $30 and Anthropic's $25, a roughly 88–90% discount, while self-reporting performance within "three to six months" of GPT-5.4 and Gemini 3.1 Pro. The model was trained with "full support" from Huawei's Ascend processors, with further price reductions expected as Huawei scales Ascend 950 production, reducing DeepSeek's dependence on Nvidia hardware. For financial institutions, the combination of near-frontier capability at a fraction of US lab pricing, open-source distribution, and a domestic Chinese chip stack materially compresses the pricing umbrella US frontier labs have relied on and sharpens questions about where inference-layer rent can be sustainably captured.
441 other tracked developments touching the same claims.