China’s AI Model Race Shifts from Token Prices to Task Costs
AGI-Signal, writing for TMTPost, argues that China’s foundation-model market is moving from token-price cuts toward competition over the cost and time required to complete real tasks. It cites October releases from StepFun, DeepSeek, MiniMax, iFlytek and Moonshot AI, including reported reductions in per-task cost, faster inference and local deployment options. Many benchmark, cost and revenue figures come from vendors or media reports and have not been independently verified.
The analysis identifies a shift in how leading Chinese model vendors differentiate their products and may affect developers’ model selection, deployment strategies and pricing negotiations.
(本文作者为 AGI-Signal,钛媒体经授权发布)

中国大模型行业,打过一场持续两年多的价格战。2024年5月,DeepSeek-V2把输入价格压到每百万Token一元,字节、阿里、百度在一周内轮番跟进,价格的一再下探此后并没有停下来,2025年2月百度宣布对话产品文心一言自4月起全面免费,3月发布的文心4.5又把API价格打到 GPT-4.5 的百分之一。
据中国信通院监测,到2025年下半年,国内大模型词元服务的归一化均价落在6.8元。需要注意的是,这一均价按模型发布时的报价归一化统计,不是用户的实际账单均价。
今年十月,这一形势似乎见了分晓。10月中下旬,阶跃、DeepSeek、MiniMax、讯飞、月之暗面在十天内接连发布新模型,五场发布里没有出现新的价格战,客户账单的变化却比过去两年都大。这中间发生了什么,要从价格战是怎么收场的说起。
#StepFun#DeepSeek#MiniMax#iFlytek#Moonshot AI#Step 5 Preview#DeepSeek V4.1 Flash-1017#MiniMax M3.1 Flash#Spark X2.6#Kimi K3.1 Pro