China's Large Model Price War: 40 Times Cheaper Than International Counterparts

The price war in China's large model market is accelerating. After the launch of V4-Pro, DeepSeek quickly reduced its API pricing to 25% of the original price, with some cached hit scenarios further reduced to one-tenth of the original price, and announced that from June 1st, the promotional pricing will be changed to long-term pricing and will not be回调ed.

Xiaomi followed suit, announcing on May 27th a permanent price cut for the series of MiMo-V2.5 large model APIs, ranging from 57% to 99% discounts, effective immediately and applicable globally.

The domestic large model subscription market is forming three pricing tiers. Entry-level products are generally priced below 40 yuan, mainly used for customer acquisition; the mainstream competitive range is concentrated between 80 yuan and 200 yuan; the high-end layer rises to 500 yuan and even above a thousand yuan, mainly targeting long context, multimodal, and Agent capabilities.

The contrast with the international market is becoming increasingly stark. Subscription fees for mainstream products like ChatGPT, Claude, and Gemini generally range from 20 to 250 US dollars, with high-end versions like Gemini Ultra and Claude Max approaching 200 US dollars. The cost gap for API usage is even more significant, with DeepSeek-V4-Pro's comprehensive API cost compressed to about 0.9 RMB per million Tokens, while GPT-5.5's long context version is about 374 RMB, Claude Opus about 204 RMB, and Gemini 3.1 Pro about 149 RMB, with price differences generally ranging from 20 to 40 times.

Market observers generally believe that this pricing gap is not a simple reflection of technological differences, but rather the result of divergence in business strategies. Companies like OpenAI, Google, and Anthropic tend to maintain high profits and enterprise-level pricing to cover high computational and R&D expenditures; Chinese model manufacturers, under competitive pressures, exchange low prices or even subsidies for developer adoption rates and Agent application penetration.

Pricing pressure is most directly reflected at the application level. Notably, in the field of video generation, there has been a contrary trend - ByteDance's Seedance 2.0, which generates a single 15-second video, saw costs increase from about 0.65 yuan to 5 yuan in the short term, a nearly sevenfold increase, indicating that compute-intensive applications are beginning to genuinely reflect the cost pressures of GPU.

This reveals the dilemma faced by the Chinese large model industry: the need for low prices to lock in developers and the ecosystem entry point, while at the same time enduring the continuous loss pressure brought about by computational, training, and inference costs. Despite revenue growth, many companies continue to report short-term losses. Analysts point out that the large model industry still needs to return to a more sustainable balance between costs and profitability in the long term.