China’s AI startup DeepSeek is moving quickly to win customers, rolling out steep price cuts immediately after unveiling its latest model, “DeepSeek V4.”
DeepSeek introduced two preview versions on April 24: the high-performance, higher-priced “V4 Pro” and the lighter, lower-cost “V4 Flash.”
According to China’s National Business Daily and other outlets, the company said April 25 it would offer V4 Pro at a 75% discount through May 5. On April 26, it also said it would cut the cost of “input cache hits” — when the same input is reused — across its product lineup to about one-tenth of the previous level.
After the change, V4 Pro’s input cache-hit fee fell to 0.025 yuan per 1 million tokens (about 5.39 won), about one-fortieth of the earlier price. Even before the discount, the pricing was lower than OpenAI’s ChatGPT 5.5 and Anthropic’s Claude Opus 4.7, the report said.
The cuts had an immediate impact. Traffic surged after launch, and on April 25 alone, calls to V4 Pro totaled 13.6 billion tokens, about four times the previous day. V4 Flash reached 50.2 billion tokens, up about 86%.
It remains unclear whether the momentum will last. V4 has not yet appeared in the weekly rankings of OpenRouter, a global AI model platform. Still, DeepSeek’s ultra-low pricing is widely seen as putting pressure on rivals. After news of the V4 discounts, shares of Chinese AI firm MiniMax fell as much as 10% intraday in Hong Kong on April 27 local time, while Zhipu AI dropped more than 3.5%.
Hu Yanping, a professor at Shanghai University of Finance and Economics, told National Business Daily that the sharp reduction in token fees was aimed at attracting large numbers of corporate and developer customers. He said it would likely pull down price expectations for competing China-based models such as Kimi, MiniMax and Qwen.
DeepSeek also upgraded performance. Both V4 Pro and V4 Flash support context windows of up to 1 million tokens, with improvements in code generation, reasoning and long-form processing. Some benchmarks have rated them as competitive with top global models. The models are described as well-suited for AI agent tasks beyond simple chatbots, though they are also known to require substantial computing resources.
Unlike earlier DeepSeek models that relied on U.S. chipmaker Nvidia, V4 was designed to run on infrastructure based on Huawei’s in-house Ascend chips.
Hong Kong’s South China Morning Post reported that the model’s release was delayed more than expected due to the shift to Huawei semiconductors, reflecting China’s push to strengthen AI self-reliance. On April 26, the social media account “Yuyuantantian,” affiliated with China Central Television, said “domestic computing power supported V4,” highlighting cooperation between DeepSeek and Huawei’s Ascend chip systems.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.