The big model price cut will not be the first round, nor the last round
A price war in AI big model industry has begun.
On May 15, at the Volcano Engine Motive Power Conference, the byte beat announced that the enterprise level price of the 32K main model of the Doubao big model was 0.0008 yuan/thousand Tokens, starting the era of "per cent". After conversion, 1 yuan can buy about 2 million Chinese characters, which is 99.3% lower than the industry price. A few days later, Alibaba and Baidu followed suit one after another. Alibaba officials announced price reductions for nine Tongyi big models, and Tongyi Qianwen GPT-4 big model Qwen Long dropped 97%. On May 21, Baidu directly announced that the two main models were free.
At present, however, Huawei and Tencent have held their ground.
In the face of the current price reduction situation, an industry related person sighed to the reporter of Time Weekly: "No way, the industry competition is so fierce, we should seize the market."
The person in charge of the volcano engine said to the reporter of Time Weekly that he welcomed the price reduction of Tongyi Qianwen big model.
Business is a cycle. Once upon a time, the smart cloud business valued by major domestic Internet companies also experienced competitive bidding. Now, the trend of price reduction is blowing from abroad. In the past year, OpenAI has made four price reductions.
Must the price of the big model be reduced? Price war has begun. Under the high cost of large models, capital bidding can sit on the card game.
Source: Ichthyozoan Price cut from overseas
As for the price reduction of domestic big factories, Lv Benfu, vice president of China National Innovation and Development Strategy Research Association, told the Times Weekly that "this price war is caused by the GPT 4.0 version."
On May 14, OpenAI released the latest multimodal large model GPT-4o at the spring new product launch and announced a price reduction of 50%. At present, the price of GPT4 is 42 cents per 1000 tokens. This is the fourth price cut of OpenAI in the past year.
In the face of GPT-4o with better performance and better price, price adjustment has also been started in China. On May 11, the start-up company Zhipu AI of Tsinghua University announced a new price system. The transfer price of the personal GLM-3 Turbo model was reduced from 5 yuan/million tokens to 1 yuan/million tokens, with a price reduction of 80%.
As a newly expensive AI company invested by Meituan, Ali, Tencent, Xiaomi and many other companies, Zhipu AI really has the confidence to fight against price. As of October 2023, Zhipu AI announced that it had accumulated more than 2.5 billion yuan of financing that year.
However, facing the high cost input of large models, large manufacturers often have more competitive advantages.
Source: Volcano Engine Propulsion Conference At the Volcano Engine Conference on May 15, Byte Beat released the bean curd AI application family, and set the enterprise level price of the bean curd Pro 32k main model at 0.0008 yuan/thousand Tokens, about 99% lower than the industry, that is, 1 yuan can buy 1.25 million Tokens of the bean curd main model, about 2 million Chinese characters.
Although the bean curd came late, the price directly "broke through the floor". For another conceptual analogy, suppose that a book of Romance of the Three Kingdoms is about 750000 words, and one yuan spent on Doubao can deal with the text volume of three books of Romance of the Three Kingdoms.
On the same day, Baidu released an announcement, in which it mentioned: "The use of the big model should not only depend on the price, but also on the comprehensive effect. Only by making the AI application effect better, the response speed faster, and the distribution channels wider, can people really feel the convenience of AI for social production." Suspected of responding to the "diving price" of bytes.
Although Baidu said that "we should not only look at the price", it soon joined the price competition. On May 21, Baidu directly announced that its two major models were free of charge.
The reporter of Time Weekly noticed from the official website that the two major models both support training optimization and deployment calls, and they are big models under the Baidu Intelligent Cloud Qianfan big model platform, while the Baidu Intelligent Cloud Qianfan big model platform is the only entrance to Wenxin big model enterprise level services.
Source: Screenshot of Baidu Intelligent Cloud official website In other words, Baidu's free and open products are mainly provided for enterprises. In addition, ERNIE Speed is a big language model just released in 2024, and ERNIE Lite is a lightweight big model.
An insider told TIME that lightweight means relatively low investment, but still high cost.
Compared with Baidu's small scale free model, Alibaba's Tongyi Qianwen big model has a wider price reduction coverage, including 9 commercial models and open source models. Among them, Qwen Long, the performance benchmarking GPT-4, has the largest price reduction. The API input price has dropped from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, a decrease of 97%; The API output price dropped from 0.02 yuan/thousand tokens to 0.002 yuan/thousand tokens, a drop of 90%.
For the input price reduction is greater than the output price, Alibaba official said that now it has become one of the most common requirements for users to ask questions about large models in combination with long texts (papers, documents, etc.), so the model input calls are often greater than the output calls.
Source: Screenshot of Tongyi Qianwen Price Reduction Announcement When the wind of price war blew, Tencent chose not to move. On May 17, Tencent held a generative AI industrial application summit. Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Cloud Intelligence, responded in a media interview that "Tencent's big model will focus on the development of products and technical capabilities in the future, and believes that it can provide competitive models and products." He did not directly respond to the outside discussion on prices.
On May 21, Alibaba and Baidu offered price trumpets one after another. The person in charge of Volcano Engine said to the reporter of the Times Weekly, "We very much welcome the price reduction of Tongyi Qianwen big model to help enterprises explore AI transformation at a lower cost and accelerate the implementation of big model application scenarios."
In addition, as for the technical strength mentioned by Baidu and Tencent, in terms of bytes, the Tokens quota per minute can reach several times of the same specification model in the industry while the price of the Doubao big model is significantly reduced, which can support a large number of concurrent requests, helping enterprises to call the big model in the production system.
"Bidding" is a cycle
When users enjoyed the price reduction effect caused by the GPT-4o version, Lv Benfu said to the reporter of the Times Weekly, "This may not be beneficial to the industry, but may lead to" clearing "."
Referring to the development process of intelligent cloud business of major Internet companies, Alibaba, Tencent and Baidu have all experienced price reshuffle. As the high expectation business of major Internet companies, Alibaba, Tencent, JD and Baidu have successively joined in the price reduction of intelligent cloud products in the past year.
In April of last year, Alibaba announced the largest price reduction in history at the 2023 Alibaba Cloud Partner Conference, and in May, it announced the maximum product discount of 40% at the 2023 Alibaba Cloud International Partner Conference. Then Tencent Cloud announced price cuts for several core cloud products, with some products falling by up to 40%. China Mobile's Mobile Cloud also announced a maximum drop of 60% for all its products.
In February this year, Alibaba Cloud launched the "Crazy Thursday", announcing that the prices of its core products were lowered across the board, with an average price reduction of 20% for more than 100 products, with a maximum reduction of 55%. On the night of Alibaba Cloud's price reduction, JD Cloud's official website released slogans such as "price comparison across the whole network, price comparison of 1 billion yuan, benchmarking products lower by 10%, and compensation if you buy more", and even directly listed the price comparison objects as "specific cloud businesses" such as Alibaba Cloud, Tencent Cloud, Huawei Cloud, etc. On April 8, Tencent Cloud even launched activities such as second killing and coupon package delivery.
Source: Screenshot of JD Cloud official account Since the establishment of intelligent cloud business by major Internet companies, the development of each intelligent cloud business has experienced a cycle of at least 10 to 20 years. When the industry is under pressure to break the ceiling of demand, price reduction has become a necessary means to obtain new users and stimulate demand growth. However, the domestic big model will enter the "Hundred Regiments War" mode in a few months in 2023, and the price reduction has been started for only one year, which means that AI big model will face more intense competitive pressure in acquiring user volume and improving demand.
From this point of view, the price cut of the big model will not be the first round, nor will it be the last round. Under the high cost input, the user usage must be increased at a faster speed to survive in the fierce competition of AI big model.
On the whole, the current major models have launched price reduction and free strategies, but the price adjustment range, open range, product type, and target users are different and have different priorities. For example, Zhipu AI has opened the personal version, Baidu is the main force of two enterprise level products, byte enterprise level products have announced a 99% price reduction, and Alibaba has selected a single model API to reduce the input price by 97%.
From the perspective of the industrial development process of AI big model, Tian Feng, president of Shangtang Intelligent Industry Research Institute, analyzed the reporter of the Times Weekly: "Data is the future currency, also the training resource of AI big model, which belongs to a new type of production factor, so the big model is priced based on the data processed, and the cheap big model is conducive to bringing national innovation in application scenarios. Engineering and infrastructure will drive the cost of the big model to continue to reduce, and when it reaches the cost inflection point, it will activate the exponential growth of social demand. "
