Big models have big price cuts! Byte, Alibaba, Baidu and Tencent take the lead in opening the book!

In the past week, major model manufacturers have announced price cuts. How come a price war has started in the blink of an eye? Why do these major manufacturers have the confidence to cut prices?

I thought the recent hot spot on the Internet should be the promotional activities of e-commerce companies, but unexpectedly, the big models of big companies stole the limelight. Within a week, ByteDance, Alibaba, Baidu and Tencent announced big price cuts! Some even offered free services! I don’t make money anymore, I just make friends and contribute to the development of the industry!

Who wouldn’t be excited after reading this news?

Let's take a quick look. On May 15, ByteDance fired the first shot in the domestic big model price war. Its AI big model Doubao (formerly known as Skylark) announced that the main model pricing is 99.3% cheaper than the industry price level! The API input price of Doubao big model is 0.0008 yuan/thousand tokens (the smallest unit of big model text). Since then, the market price of Chinese big models has directly entered the "centimeter era". 1 yuan can buy a full 1.25 million tokens at Doubao!

ByteDance may be worried that friends at the scene would not have an intuitive feeling, so it specially compared the large models of two friendly competitors, Wenxin Yiyan and Tongyi Qianwen. It was indeed simple and clear, and shocking!

Image credit: Volcano Engine

ByteDance's Doubao was once in the limelight, but those who were compared with it were a little restless. Price war? Who wouldn't! On May 21, Alibaba Cloud announced that Tongyi Qianwen also had a big price cut, and the magnitude was even greater! Tongyi Qianwen was directly reduced to 0.0005 yuan/thousand tokens, and 1 yuan could buy 2 million tokens, which is equivalent to 5 Xinhua dictionaries, and announced in advance that it won the title of "King of Cost-Effectiveness".

Okay, okay, that’s how we play it, right?

In the afternoon, Baidu, which is the "big brother" in the domestic large model market, released an announcement on its official public account with only one sentence: "The two main models of Wenxin Large Model are completely free, effective immediately!"

Image source: Baidu Smart Cloud

The shorter the news, the bigger the news. The sentence that the main model is free is more shocking than the "floor price". On the 22nd, Tencent could no longer hold back and announced a new large model upgrade plan. The Hunyuan-lite model, one of the main models, was not only upgraded from 4k to 256k, but also completely free; in addition to the length upgrade, the prices of other models have also dropped significantly.

Image source: Tencent

After reading all the news about price cuts by major domestic manufacturers, I just want to say: "It's brutal, too brutal!"

But for users, it is a great news, as the cost of using large models has been greatly reduced. We have also sorted out the latest large model prices of four major manufacturers, so that everyone can grasp the first-hand discount information.

First, there are the newly added free models. I would like to specifically praise Baidu and Tencent. Alibaba Cloud Tongyi Qianwen also has two open source models that are free for a limited time of seven days.

Next, let's compare the prices of the main models of major manufacturers. (Note: Large language models are usually charged separately for input and output. This is because the input and output resource consumption of the model are different during the inference process. Input billing is based on the request data submitted by the user to the model, and output billing is based on the output results returned by the model to the user.)

Since different models have different parameter quantities and maximum lengths of supported tokens, for example, Alibaba's Qwen-Long has a maximum context length of 10 million, while Baidu's ERNIE-4.0-8K is still 8k, it would be unfair to only look at the price when making a horizontal comparison between major manufacturers. Therefore, we have also made the price of each manufacturer's model for reference.

Not long ago, we were still complaining about large models being too expensive, but how come a price war has broken out in the blink of an eye?

In my opinion, the successive price cuts by the four major manufacturers reflect that the domestic large-scale model industry is rapidly entering the stage of large-scale commercialization. Lower large-scale model use costs are the prerequisite for development at this stage. If large models are too expensive, how can enterprises implement them? Without implementation, how can the industry develop? The price cuts by the major manufacturers mean much more to "make friends" than to make money. With more friends, the large-scale model ecosystem of the company will prosper, which is the real future to look forward to.

On this point, Liu Weiguang, senior vice president of Alibaba Cloud Intelligence Group and president of the Public Cloud Business Unit, made it clear when announcing the price reduction that "the purpose (of the price reduction) must be to benefit the market" and "to truly accelerate the early outbreak of the market."

And with the advancement of technology, just like Moore's Law, the cost of large models is also continuously decreasing. In the early morning of May 22, at the Microsoft Build Developer Conference, Nadella said that the performance of GPT-4 has increased by 6 times in the past year, but the cost has been reduced to 1/12 of the previous level, corresponding to a 70-fold increase in performance/cost.

In the previous stage of the development of large models, the competition was about technical level, and ranking evaluation was the mainstream; in this stage, comprehensive strength has become more important. However, in the domestic market, the battle of large models has been brought into a white-hot state by large manufacturers. Speaking of this, it is easy to think of a classic line:

Image source: Internet

To be in the big model race, you also need influence and background. Liu Weiguang believes that not everyone is qualified to cut prices. In China, companies that are qualified and capable of cutting prices (for big models) should meet the following four conditions:

1. The model's basic model capabilities are sufficiently advanced and have truly comprehensive multi-modal capabilities;

2. There are real inference resources, including the support of basic data center capabilities such as network, basic computing, and storage behind large-scale inference computing clusters;

3. The current model is already used by many customers and is producing commercial results;

4. Large models are the company’s core business.

It is not difficult to find that these four large companies are confident in reducing prices and have the ability to accelerate the implementation of large models. The collective price reduction of large companies may be a bit cruel to large model startups, but it is a big boost to the industry. Liu Weiguang said that he believes that the daily call volume of China's large models will soar from 100 million to 10 billion in 2024. With this round of "price reduction promotions", this prediction may really come true.