Chinese artificial intelligence startup DeepSeek said it will permanently maintain steep discounts on its DeepSeek-V4-Pro model API pricing, pushing inference costs to new industry lows and escalating a price war across the global AI sector.

DeepSeek said the promotional pricing for its V4-Pro model, previously offered at 25 percent of the standard rate and originally scheduled to end on Sunday, would now become permanent. The company said the revised pricing effectively sets the model's official price at one quarter of the originally planned level.

Under the new pricing structure, input costs for cached requests will fall to 0.025 yuan ($0.0037) per million tokens, while input costs will be 3 yuan per million tokens and output costs 6 yuan per million tokens, according to the company. The pricing ranks among the lowest globally for mainstream large-language-model APIs.

The move comes as AI infrastructure costs are rising worldwide due to what industry experts describe as structural imbalances across the AI supply chain.

DeepSeek's decision to cut prices against that backdrop signals that competition in China's AI market is increasingly shifting from raw computing scale toward efficiency and ecosystem expansion, industry experts said.

Wang Peng, a researcher at the Beijing Academy of Social Sciences, said: "DeepSeek's willingness to cut prices against the market trend is absolutely not simple cash-burning subsidization. It is a cost advantage achieved through reconstruction of the underlying technology architecture."

Wang said DeepSeek had adapted its models to domestic Chinese computing platforms including Huawei's Ascend chips, reducing dependence on overseas high-end computing hardware and lowering procurement costs.

"It will create a low price-user growth-ecosystem prosperity, which will further decline cost and form a cycle that strengthens DeepSeek's competitive position," he added.

The move also reflects a broader shift in China's AI industry, where firms are increasingly attempting to differentiate themselves not only through model capability but through deployment efficiency, infrastructure optimization and lower-cost inference.