The domestic large model industry is entering a turning point in pricing logic.
On August 6th, DeepSeek announced that it plans to increase the overall pricing of its DeepSeek API services in the near future, with an expected significant increase. The specific plan will be subject to official notification. Although DeepSeek has not yet announced the magnitude and timeline of the price adjustment, the market generally believes that this means that the domestic large model industry is bidding farewell to the previous stage of seizing the market with low prices and entering a new cycle centered on cost, efficiency, and commercialization.
Behind this price increase is the dual pressure brought by the explosive demand for AI applications and the rising cost of computing power. With the rise of complex applications such as AI agents, the number of model calls has significantly increased, and the model that relied on low prices to stimulate demand in the past is facing challenges.
The surge in usage has forced DeepSeek to reprice its pricing
The core reason for this price increase is that the growth rate of demand for AI applications has exceeded expectations.
According to OpenCode, an open-source AI agent tool, the DeepSeek V4 Flash official version has a daily token call volume of 8 trillion on its platform, of which 5 trillion comes from free quotas and 3 trillion comes from paid packages. As a comparison, OpenRouter, a large model routing platform, has access to over 400 models, with a daily average token call scale of approximately 6.6 trillion across the entire platform.
This means that the number of calls of DeepSeek single model on a single platform has exceeded the average daily scale of the whole model aggregation platform.
According to data from the Vercel platform, DeepSeek Token processing continues to climb, with V4 Flash processing approximately 5.3 trillion tokens per week, making it one of the most widely used models on the platform.
With the rapid popularization of AI agents, the model calling pattern is undergoing changes. In the past, chatbots could only consume hundreds to thousands of tokens for a single interaction, while when agents execute complex tasks, they need to continuously call models and access tools, and token consumption may reach tens or even hundreds of times that of traditional conversation modes.
Under the explosion of demand, extremely low prices are becoming a new cost pressure.
DeepSeek previously stated that the price increase is mainly influenced by three factors: firstly, the rapid growth of intelligent agent applications leading to a surge in token consumption; Secondly, the supply of high-end AI chips is limited, and the cost of computing power continues to rise; Thirdly, industry competition is shifting from subsidy for scale to sustainable business models.
From 'Price Butcher' to Commercial Pricing
The pricing shift of DeepSeek did not occur suddenly.
At the end of April this year, DeepSeek launched a limited time discount on the V4 Pro API, and then switched the discounted price to long-term pricing in May, resulting in a 75% decrease in the output token price compared to the original price. At that time, the output price of DeepSeek V4 Pro was about 0.87 US dollars per million tokens, far lower than mainstream models overseas, and was known as the "price butcher" by the market.
This strategy has also triggered a collective price reduction among major domestic model manufacturers, with Xiaomi, ByteDance, Tencent Cloud, and others adjusting their pricing systems one after another. Some model prices have dropped by more than 90%, and the domestic large model market has been plunged into fierce price competition.
But at the end of June, DeepSeek began to change its strategy. The company announced the introduction of peak valley pricing mechanism after the official launch of V4, doubling the API price during peak hours on weekdays. At that time, the market believed that this was DeepSeek's way of adjusting user demand and optimizing computing power utilization through pricing.
And this comprehensive price increase means that DeepSeek is shifting from "regulating traffic" to "redefining prices".
DeepSeek founder Liang Wenfeng previously stated at an investor conference that API pricing should be based on reasonable recovery of device costs, with the goal of "buying a batch of devices and recovering costs within ten months". He also pointed out that there is "almost no elasticity" in user demand within the current price range, and even if prices rise, changes in token consumption are limited.
Shortage of computing power and difficulty in sustaining low price models
Behind DeepSeek's price adjustment is the intensifying supply-demand imbalance in the entire AI computing power industry chain.
With the rapid growth of training and inference requirements for large models, GPU、 The resources such as servers and storage continue to be tight. Especially in the era of AI agents, the demand for model inference has shifted from simple question answering to long link task execution, resulting in a significant increase in computing power consumption.
Goldman Sachs previously pointed out that DeepSeek's implementation of peak valley pricing does not represent weak demand, but rather reflects the strong demand for AI models and the tightening of computing resources in China. As AI applications enter the stage of scale, maintaining extremely low API prices in the long run will be difficult to cover the continuous growth of computing power investment.
In fact, in the past six months, there have been signs of collective adjustment in the domestic large model industry.
Many manufacturers, including Alibaba, ByteDance, Xiaomi, Zhipu, Tencent, etc., have adjusted their product pricing or package systems, and some low-priced solutions have begun to shrink. The API call price of Zhipu increased in the first quarter of this year compared to before, but the call volume still maintained high-speed growth; Kimi also suspended some new user subscriptions due to significantly higher than expected user request volumes.
The industry is shifting from "burning money for users" to "large-scale commercial monetization".
Domestic large models enter profit verification period
DeepSeek's price increase this time also sends a more important signal: domestic large models are transitioning from technological competition to commercial competition.
In the past year, low-priced or even free model services have helped domestic large models rapidly expand their user base. However, with the explosion of call volume, the cost of computing power, server investment, and operational pressure have begun to emerge.
The market previously focused on who could provide cheaper models; In the future, the focus of competition will shift towards who can provide higher performance and more stable services at lower costs.
For developers, the price increase of DeepSeek means that the cost dividend of AI applications is decreasing, but it also means that the industry is moving towards a healthier business cycle.
