Recently, the domestic large model company DeepSeek announced a major decision: the API price of its flagship model DeepSeek-V4-Pro will be permanently reduced to 25% of the original price. According to the official announcement, after the limited-time 25% discount expires on May 31, 2026, this price will officially take effect and remain permanently. The specific pricing is 0.025 yuan per million Tokens input (cache hit), 3 yuan per input (cache miss), and 6 yuan per output. This price reduction has brought the API price of this company to a new low in the global large model field.
This permanent price reduction initiative by DeepSeek occurred against the backdrop where the global AI large model industry is generally facing cost pressures, and some cloud service providers have raised the prices of their APIs. Since 2026, the global AI large model industry has been mired in a price increase trend. The price of HBM (high-bandwidth memory) has soared by over 500% in half a year, high-end GPUs (graphics processors) are in short supply, and the demand for inference-end tokens (tokens) has skyrocketed. Amazon, Microsoft, and major domestic cloud providers have all raised the pricing of APIs (application programming interfaces), with some increases even reaching as high as 463%. The industry seems to have reached a consensus that "AI services should become increasingly expensive".
DeepSeek released its new-generation flagship large model, DeepSeek-V4, and made it open-source on April 24, 2026. Official information indicates that compared to the previous generation model, the Agent capability of DeepSeek-V4-Pro has significantly improved. It achieved the best level among current open-source models in the Agentic Coding evaluation. In multiple assessments such as world knowledge, mathematics, STEM, and competitive code, its performance surpassed all publicly evaluated open-source models and was comparable to the results of the world's top closed-source models.
So, why did DeepSeek dare to cut prices in a declining market? Some analyses suggest that this move is not about "spending money to gain market share", but rather stems from the company's ability to control costs through the reconfiguration of its underlying technologies. The key technological breakthroughs are mainly reflected in three aspects:
l Firstly, there is an innovation in the architecture, with the independently developed sparse attention mechanism and hybrid expert model. This enables the V4 series to consume only 27% of the computing power compared to previous versions when processing long-context with millions of tokens. The key-value cache occupancy is reduced to 10%, thus achieving a technical reduction in unit inference cost.
l Secondly, there is autonomy in the computing power aspect. By deeply adapting to domestic computing power platforms such as Tsinghua Unigroup, it reduces reliance on overseas high-end computing power and lowers hardware procurement costs.
l Thirdly, there is continuous optimization on the engineering side. In the inference process, extreme efficiency is enhanced, improving the computing power utilization rate, and through the scale effect, fixed costs are shared, forming a virtuous cycle of "increased usage volume compensating for cost reduction". This cost reduction driven by technological progress provides a sustainable foundation for price adjustments.
Almost at the same time, according to Bloomberg, DeepSeek is in the process of launching a financing round with a scale of up to 70 billion yuan (approximately 10 billion US dollars). If it is ultimately completed, this would likely set a new record for the largest single financing deal in the history of Chinese AI enterprises. Sources close to the matter disclosed that the pre-investment valuation of DeepSeek in this round might be around 45 billion US dollars. Potential investors include the National Artificial Intelligence Industry Investment Fund, Tencent, IDG Capital, Li Si Capital, among others. NetEase and JD.com are also in discussions about joining.
It is reported that Liang Wenfeng, the founder of DeepSeek, stated at an investor meeting that the company will continue to prioritize breakthrough artificial intelligence research and strive to achieve general artificial intelligence (AGI), while also committing to continuing the development of open-source AI models. Liang Wenfeng himself may have personally invested approximately 20 billion yuan in this round of financing.
