Xiaomi MiMo V2.5 API cuts spark 99% China AI price war

Xiaomi’s 99% API price cut is a clear signal that China’s leading labs are willing to trade short‑term margin for API volume, data and ecosystem lock‑in. By pushing MiMo‑V2.5‑Pro cache‑hit pricing down to fractions of a cent per million tokens—explicitly matching or undercutting DeepSeek’s V4 pricing—Xiaomi is trying to make it irrational for domestic developers to pay Western rates for GPT‑5‑series or Claude tokens. The article makes it clear this isn’t pure subsidy: Xiaomi claims major engineering wins in cache efficiency and token reuse, letting it lower effective compute per request.([m.voc.com.cn](https://m.voc.com.cn/xhn/news/202605/32770665.html))

In the race to AGI, a sustained price war matters because it shifts the constraint from token cost to imagination. If frontier‑ish Chinese models are 10–50x cheaper per unit of useful work than US counterparts, you’ll see many more always‑on agents, longer context windows and more experimental workflows running on Chinese infrastructure. That accelerates learning loops for Chinese labs and increases their leverage in model‑distillation battles.

At the same time, the piece notes that many other domestic providers—including Alibaba and Tencent—have been raising prices under chip‑supply pressure. Xiaomi and DeepSeek are betting they can win by out‑optimizing on inference efficiency and by leaning on domestic accelerators like Huawei Ascend, a dynamic that reduces Nvidia leverage and makes the compute race more multipolar.