DeepSeek’s New Launch! API Prices Slashed by Over 50%

The staff led by DeepSeek founder Wenfeng Liang has launched a brand new experimental mannequin, DeepSeek-V3.2-Exp, on September 29, marking a big step in exploring next-generation transformer architectures. Launched as an open-source transitional product, its core improve lies within the introduction of DeepSeek’s proprietary Sparse Consideration (DSA) mechanism, designed to optimize coaching and inference effectivity for long-text processing.

Wu Chao, Chief TMT Analyst at CITIC Securities, commented that the brand new model “considerably enhances usability.” Technically, the DeepSeek DSA mechanism achieves fine-grained sparse consideration for the primary time. In public benchmark assessments throughout numerous domains, its output high quality matches that of the earlier V3.1-Terminus whereas considerably bettering long-text processing effectivity, because of rigorous alignment validation in coaching configurations.

The discharge’s spotlight is its contribution to the open-source ecosystem. Along with the usual NVIDIA CUDA model, DeepSeek has open-sourced a TileLang model of GPU operators. Developed by Peking College’s Yang Zhi staff, this programming language compresses FlashAttention operator code from over 500 traces to simply 80, sustaining efficiency whereas offering builders with a user-friendly debugging instrument. Main platforms like Huawei Ascend and Cambricon have accomplished mannequin variations, with open-source inference code additionally launched concurrently.

Excitingly for builders, API costs have been drastically lowered: the enter cache hit value dropped from 0.5 CNY per million tokens to 0.2 CNY, the non-cache hit value fell from 4 CNY to 2 CNY, and the output value was halved from 12 CNY to three CNY, leading to an general price discount of over 50%. The official app, internet platform, and mini-program have all been up to date to mirror these modifications.

This launch coincides with a wave of home large-scale mannequin iterations. On the 2025 Cloud Habitat Convention, Alibaba Cloud unveiled seven new merchandise, with its flagship mannequin Qwen3-Max leveraging 36 trillion information factors and trillions of parameters to boost programming and agent capabilities. Zhipu’s GLM-4.6 is ready to debut quickly, whereas Moonshot AI’s Kimi has begun beta testing its “OK Pc” Agent mode. Business competitors is more and more centered on effectivity and ecosystem improvement.