On September 29, 2025, DeepSeek made an announcement that its official app, web - based version, and mini - program had all been simultaneously upgraded to the DeepSeek - V3.2 - Exp version. Alongside this upgrade, there was a substantial cut in API prices. In scenarios where cache hits occur, the input price per million tokens has plummeted from 0.5 yuan to 0.2 yuan. When the cache fails to hit, the input price per million tokens has dropped from 4 yuan to 2 yuan. Moreover, the output price per million tokens has decreased from 12 yuan to 3 yuan. According to the new pricing scheme, developers will see their API - calling costs slashed by more than 50%. This latest version features the self - developed DeepSeek Sparse Attention mechanism. This mechanism enhances the efficiency of long - text training and inference without compromising the performance, which remains on par with that of the previous model.