Analysts at Bank of America report a structural shift in the Chinese AI market. Providers are moving away from flat pricing per token toward tiered models focused on the total cost of task completion. This change follows the release of low-cost models by DeepSeek, which significantly reduced market prices for developers.
According to the Silicon Data index, the average inference price reached a yearly low of $1.16–1.18 per million tokens in August, down from $2.04 in late May. Additionally, some providers have introduced dynamic pricing; for instance, the V4 model rate doubles during Beijing peak hours, rising from 6 to 12 CNY per million output tokens.
Experts note that token-based pricing no longer reflects actual customer costs, especially for agentic workflows involving multiple model calls, tool usage, and human validation. Consequently, ML teams are shifting their procurement metrics to prioritize the total cost of achieving a reliable result, factoring in success rates and overhead expenses.
Premium pricing is now reserved for models that offer advanced reasoning, long context windows, multimodality, and agentic capabilities. While basic inference is becoming a commodity, these high-end features allow providers to maintain differentiation in a market that is moving beyond simple price wars.
