Article | Zhang Jun of Sina Technology

There was some excitement in the tech world yesterday.

It was thought that the e-commerce platforms were going to stage a battle between the dragon and the tiger on 618, but it was the other giants that really started the war-in just three hours, byte beat, Aliyun, Baidu and so on successively launched a "price war" on the big model.

The first one to make the move is bytes.Solcasinonodepositbonuscodes2022Claim that the pricing of the big model has been reduced to 0Solcasinonodepositbonuscodes2022. 0008 yuan / thousand tokens!

Ali Yun is not far behind: 0.0005 yuan / thousand tokens!

Baidu is to lift the table directly: the two main models are all free!

.

Can the price war of large models promote the development of the industry?

At present, it is divided into pros and cons: one side thinks that it is a good thing, which lowers the threshold and helps to speed up the application of the big model; the other side thinks that it is a bad thing, the price war will compress the profit margin, and the lack of funds will not attract more talents. if it hinders technological progress, the evolution of the model will be directly damaged, which will be more disadvantageous in the long run, and even a number of small and medium-sized start-ups will be eliminated.

The giant set off a price war: he reduced the price! You reduce the price! I'm free!

In fact, before this, the domestic large model enterprises appeared the behavior of price reduction. For example, the price of DeepSeek-V2 is 1 yuan per million tokens, and Zhisu AI has also adjusted the price of glm3-turbo to 1 million tokens1 yuan.

What triggered this wave of large-scale price war was the byte-beating entry.

A week ago, byte beat held a big model strategy conference, the most striking of which was its radical pricing strategy. Byte jump announced that the bean bag model will officially start external service. Tan Dai, president of the volcano engine, said: "the bean bag main model is priced at only 0.0008 yuan per thousand tokens in the enterprise market. 0.8% can handle more than 1500 Chinese characters, which is 99.3% cheaper than the industry." He believes that this will help enterprises to use large models at a very low cost and accelerate the landing of large models. "Technology-driven extreme performance-to-price ratio can really create value."

One stone stirs up thousands of waves.

May 21, Aliyun announced Tongyi Qianqian GPT-4 main model Qwen-Turbo, Qwen-Plus, Qwen-Long, Qwen-Max and other nine major AI models announced a comprehensive price reduction. Among them, the input price of Qwen-Long API dropped from 0.02 yuan / thousand tokens to 0.0005 yuan / thousand tokens, with a maximum drop of 97 percent.

After the comparison of Sina Science and Technology, it is found that after this price reduction, the price of Tongyi Qianwen is 0.0005 yuan / thousand tokens, while the price of the previous bean bag model is 0.0008 yuan / thousand tokens. This means that the price of Tongyi Qianwen is cheaper than the big bean bag model under the same number of tokens.

To make the smell of gunpowder stronger, the person in charge of the volcano engine responded directly to the competitor's price reduction: Tongyi Qianqian model price reduction is very welcome, together to help enterprises to explore AI transformation at a lower cost and accelerate the landing of large model application scenarios.

However, the price war between the giants is not over. On the afternoon of May 21, ERNIE Speed and ERNIE Lite, the two main models of Baidu Wenxin model, announced that they were all free of charge, effective immediately. Baidu this action sniper byte jump and Ali Yun's intention is very obvious, but also introduced this wave of price war into the most exciting part.

Support: reduce costs and promote bundling

For the price war of the big model, the industry also has different views. There are entrepreneurs and people in the industry who support the price war.

Kaifu Lee, chairman and chief executive of Innovation works, told Sina Technology that to a certain extent, the entire industry reduces the reasoning cost by 10 times every year, which can be expected and should happen. "this is good news and good news for the industry as a whole."

An industry source told Sina Technology that the big model industry has indeed reached the stage of pursuing landing and measuring commercialization, so everyone begins to consider using more aggressive price reduction measures to get customers, and a price war is bound to happen. At the same time, he pointed out that at present, most of the enterprises that cut prices are cloud manufacturers, or large model companies with cloud capabilities, such as Doubao backed by volcano engine, Wenxin big model backed by Baidu Smart Cloud, Tongyi Qianqian asking about Aliyun, they are not all about price reduction, but only a small part of them have reduced prices to gain market attention through this partial marketing color way. After attracting customers, they have a larger volume and more services, and they can have some bundled sales behavior, so they can make profits as a whole.

He believes that small and medium-sized start-up large model enterprises will certainly suffer some impact in the price war. The industry has been rolled up, and after all, some enterprises will die. "now the competition is not necessarily whose technology is strong, technology may be one of the factors, but also compared to whose financing ability is stronger, whose business ability is stronger, and finally can really stand."

Opposition: vicious circle, widening gap with first-class level

Many experts have also expressed concern about this.

Jia Yangqing, founder of Lepton AI and former vice president of Alibaba, pointed out that from the point of view of the entire AI industry, price reduction is a simple strategy that can be done with a pat on the head, but the real To B business success is more difficult. "Today, it's not that API is expensive and nobody uses it, but because companies have to figure out exactly how to use it to generate business value, otherwise, no matter how cheap it is, it will be a waste, and today's implementation of this layer is missing." Jia Yangqing said, "perhaps not the cheapest to win the business war, but the best way to win profits on the ground."

There is also a view that if the model ability of domestic manufacturers can not reach the level of OpenAI, blindly fighting prices is not a good strategy. For example, OpenAI recently announced the launch of GPT-4o, an upgraded version of GPT-4, which enables multi-modal interaction of text, audio and images. Google also showed a full range of AI cards at the Igamo developer conference, deeply integrating AI technology into its huge product ecology. Once the domestic large model enterprises fall into a simple price war, the cash flow will continue to be under pressure, and there is a lack of sufficient funds to invest in recruiting AI talents and upgrading infrastructure, which will also affect the evolution and iteration of the model, thus making the domestic base model enter a bottleneck period of development, and the gap between the domestic base model and the world-class level is widening.

It is worth noting that at the same time of the price war for large models, Wu Jiangxing, academician of the Chinese Academy of Engineering, director of the National Digital switching system Engineering Technology Research Center, and director of the big data Research Institute of Fudan University, recently pointed out that "the endogenous security of AI is the biggest obstacle to the current application and promotion of AI." He said that a security analysis of the 10 most popular large models found that more than 90% of the mainstream large models were unreliable.

For domestic large model companies, in addition to price wars, a strategy that attracts attention, they also need to make great efforts in terms of technical level, safety and other aspects to improve their comprehensive capabilities, so as to promote the healthy and healthy development of the industry.

: Jiang Yuhan