A 75% discount on DeepSeek-V4-Pro will be available to developers through May 5, the Chinese AI startup announced, while a separate change takes the price of input cache hits across its full API lineup down to one-tenth of what they had been — a reduction that takes effect right away.
At full price, before the promotional discount, V4-Pro costs $0.145 per million input tokens and $3.48 per million output tokens, according to The Next Web. After the promotional discount is applied, input tokens on V4-Pro work out to roughly $0.036 per million. The lighter V4-Flash option carries a standard-rate price of $0.14 per million input tokens and $0.28 per million output tokens. Even at standard pricing, both models undercut comparable offerings from OpenAI, Google $GOOGL, and Anthropic, according to The Next Web.
Cutting the cache-hit rate to a tenth of previous levels is aimed squarely at the segment of users whose workloads involve high volumes of repeated or near-identical queries — a pattern that characterizes most real-world AI agent deployments, according to The Next Web.
DeepSeek launched a preview of V4 last week. V4-Pro uses a mixture-of-experts architecture housing 1.6 trillion parameters in total, with 49 billion active at any given time; V4-Flash, the smaller sibling, comes in at 284 billion total parameters and 13 billion active. Both support a one-million-token context window. According to Reuters, DeepSeek claims V4-Pro leads every open-source rival on world-knowledge evaluations, with only Google's proprietary Gemini-3.1-Pro scoring higher. The model has also been optimized for AI agent tasks and integrates with tools including Anthropic's Claude Code, according to Reuters.
The price cuts arrive against a charged geopolitical backdrop. Michael Kratsios, the White House's Director of Science and Technology Policy, made a pointed accusation last week: foreign actors — with China at the center of his concern — have been systematically copying U.S. frontier models by feeding a larger model's outputs into a smaller one, allowing the smaller model to absorb comparable capabilities without the underlying research investment. His memo stopped short of identifying DeepSeek by name, though the Hangzhou firm has separately faced public accusations of model distillation from both OpenAI and Anthropic, according to CNBC.
The release of R1 in January 2025 sent a jolt through the AI industry after DeepSeek disclosed that the reasoning model had been produced in roughly two months at a cost below $6 million, relying on chips less powerful than those typically used by U.S. labs — figures that led many to question whether American tech giants were over-investing in hardware. In the period that followed, rivals such as Alibaba and ByteDance moved quickly to bring out their own models, sharpening the competitive landscape inside China. DeepSeek's V4 model is also adapted to run on Huawei's Ascend AI processors, a shift away from the company's prior reliance on Nvidia $NVDA hardware.
The broader competitive pressure from Chinese open-source AI has been building steadily. Chinese open-weight models have become deeply embedded in the U.S. AI ecosystem, appearing in developer tooling, cloud marketplaces, and production applications — with some estimates putting Chinese open-source models inside about 80% of U.S. AI startups. U.S. cloud providers including Amazon $AMZN, Microsoft $MSFT, and Google have all made Chinese models available on their platforms. The pricing gap has been a central driver: one analysis found Chinese models can run at one-sixth to one-fourth the cost of U.S. rivals.