DeepSeek, Xiaomi Just Made Frontier AI 99% Cheaper. American Labs Went the Other Way

The economics of frontier AI are diverging. While American labs are raising the cost of intelligence, Chinese labs are aggressively driving it toward zero. Recent price shifts from DeepSeek and Xiaomi show a massive gap in how the next generation of AI-powered products will be priced.

DeepSeek has made its 75% V4-Pro discount permanent, locking in output at $0.87 per million tokens. Simultaneously, Xiaomi slashed MiMo-V2.5 API prices by up to 99% for cached inputs, bringing the cost for the Pro model down to $0.0036 per million tokens.

This is not just a marketing tactic; it is a structural shift in inference efficiency. Xiaomi’s system can now remember about five times more data than before, reducing storage and processing costs by around 80%. The impact on scale is immediate: Xiaomi's Max plan at $100 now provides 82 billion tokens, an increase from 1.6 billion. To put that in perspective, 82 billion tokens represents more than 60 billion words.

The contrast with American labs is stark. OpenAI's GPT-5.5 doubled output prices to $30 per million tokens at launch. Meanwhile, Anthropic's Claude Opus 4.7 introduced an updated tokenizer that can inflate actual costs by up to 35%.

This divergence creates a massive strategic divide for developers. When building an agent or a web site, token pricing determines whether a product is economically viable or a money pit. If you are building high-volume, automated workflows, the cost of intelligence is no longer a fixed overhead—it is a variable that can be optimized by choosing the right architecture.

The winners in this cycle will be the builders who move away from high-margin, high-cost models toward these highly efficient, low-cost alternatives. The question for your roadmap is whether you are building on a foundation of rising costs or falling ones.