China's brutal AI price war may be maturing into something more familiar: tiers. On July 19 at WAIC, executives signaled a shift away from the uniform low pricing that has defined the domestic market, toward deliberate segmentation of customers by willingness to pay.
The case for cheap
MiniMax architect Bai Chuanxu defended the aggressive low end, pointing to how far capability-per-dollar has moved: "GPT-3.5 cost $20 a month back in the day. Today's domestic 40-yuan (about $5.50) subscription models are generally much stronger than that." In that view, cheap tiers are not loss-leaders but a reflection of collapsing inference costs.
The case against it
Alibaba Cloud representatives pushed back, arguing that 40 yuan is unsustainable and that the 200-yuan (about $27.50) tier has better long-term prospects. The disagreement is really about whether the market rewards volume at razor-thin margins or premium plans that can actually fund frontier training.
Free at the bottom
iFlytek illustrated the third layer, unveiling a desktop agent product with a daily free quota for light users. Stacked together, the announcements sketch a market splitting three ways: free tiers to acquire users, cheap mass plans, and pricier premium and pay-per-token options for heavy demand — the same structure that eventually stabilized Western AI pricing.
Why the shift now
Segmentation is what happens when a price war stops being winnable. Uniform low pricing is a customer-acquisition tactic; it does not pay for the next model. The move toward tiers suggests Chinese labs are beginning to prioritize monetization over raw share — a maturation that echoes how the pricing debate has played out globally, including Anthropic's own move to meter heavy users.