DeepSeek Announces Nighttime API Call Price Reduction, with Discounts Up to 75%
On February 26, DeepSeek announced that it will significantly reduce the price of API calls during the night idle period from 00:30 to 08:30 Beijing time, with DeepSeek-V3 dropping to 50% of the original price, and DeepSeek-R1 even lower to 25%, with the highest reduction reaching 75%. The company said that this move aims to encourage users to fully utilize the night time period and enjoy a more economical and smoother service experience.
This is not the first time DeepSeek has adjusted its API prices. In August of last year, DeepSeek once reduced the API input cost to 0.1 yuan per million tokens and the output cost to 2 yuan per million tokens, causing a stir in the industry.
However, in early February, due to server resource constraints, DeepSeek temporarily suspended API service recharging until February 25, when it was reopened and the price of the DeepSeek-chat model was adjusted to 2 yuan per million input tokens and 8 yuan per million output tokens.
Comparing the call prices with those of DouBao and TongYi QianWen, one can see that under normal hours, the price of DouBao’s large model is relatively lower, TongYi QianWen’s different model prices vary, while DeepSeek’s price is relatively higher during standard hours, but there is a significant drop during the discount period, making it competitive.
Referring to DeepSeek’s previous call prices, within the discount period, DeepSeek-V3’s input (cache hit) per million tokens is 0.25 yuan, input (cache miss) is 1 yuan, and output is 4 yuan; DeepSeek-R1’s input (cache hit) per million tokens is 0.25 yuan, input (cache miss) is 1 yuan, and output is 4 yuan.
DeepSeek’s price reduction strategy helps to further lower the threshold for small and medium-sized enterprises and individual developers, promoting the popularity of AI technology. Previously, DeepSeek’s price reduction had triggered a chain reaction in the industry, with leading companies including Alibaba Cloud, Baidu, and Tencent following suit and announcing their own price reduction measures. Whether this night discount will once again trigger a price war depends on the further actions of other large model companies.
DeepSeek launched the discount measures at the time of its “Open Source Week”. The first open-source code library was FlashMLA, optimized for Hopper GPUs, the second day was the first open-source EP communication library DeepEP for MoE model training and inference, and on the 26th, an FP8 GEMM library that supports dense and MoE GEMM was open-sourced, providing support for V3/R1 training and inference.
There have also been reports that DeepSeek plans to release its new generation AI model R2 ahead of schedule, which is expected to improve in programming capabilities, multi-language inference, and possibly upgrade in multimodal areas such as image generation, voice recognition, and synthesis. The release of the R2 model may further consolidate DeepSeek’s position in the AI field and drive a new round of technological innovation.
