Price reduction! Free! Enclosure! The price war of large models is in full swing

This article mainly describes a series of price reduction and free strategies launched by major companies in the field of artificial intelligence in the large model market, aiming to attract more users and market share by lowering prices, thereby seizing the initiative and promoting their own technological innovation and development.

Giants are coming out one after another! Price cuts! Free! The price war in the large model track has gone completely crazy.

Recently, Baidu and Alibaba announced price adjustments for their large models on the same day, with Alibaba reducing prices by as much as 97%, while Baidu was even more exaggerated, directly announcing that it would be free.

Judging from the timeline, Baidu and Alibaba’s adjustment of large model prices is more of a “following the trend”. From DeepSeek, Zhipu AI, to ByteDance and Tencent, these large model companies’ successive price cuts and benchmarking behaviors have forced Baidu and Alibaba to respond.

However, on the other side of the fierce battle, all the large model companies participating in the battle have the same idea.

By lowering prices, we can enable more companies and consumers to use large-model technology, quickly expand the market share, continuously optimize technical capabilities, and gradually break out of the commercialization dilemma.

Burning money has always been the most common strategy in the Internet field. It is suitable for the large model industry and can make this market larger. However, its disadvantages of slow technological breakthroughs, accelerated industry reshuffle, and the formation of oligopoly will also be highlighted in the large model industry.

Whether players can break away from the crude "burning money" mentality and increase technology while increasing prices will determine the future of players and the direction of the entire industry.

1. Price reduction is free and effective immediately

The player who kicked off the domestic large-model price war is DeepSeek, which is known as the "Pinduoduo" in the AI ​​industry.

On May 6, DeepSeek reduced the price of its large model DeepSeek-V2, which is comparable to GPT-4, to 1 yuan per million token input and 2 yuan per million token output (32K context).

The price, which is less than one percent of GPT-4, has caused a sensation in the large model market and also aroused the "competitiveness" of other players.

A week later, the domestic star startup team Zhipu AI responded to the challenge by reducing the price of its entry-level large model GLM-3-Turbo (context length 128k) from 0.005 yuan/thousand tokens to 0.001 yuan/thousand tokens. In addition, the tokens given to new registered users of the open platform increased from 5 million to 25 million (including 20 million GLM3-Turbo and 5 million GLM4).

The actions of the second-tier players have also aroused the interest of major companies.

Four days after Zhipu AI lowered its price, ByteDance's Volcano Engine suddenly launched a surprise attack and announced the lowest price in history for its Doubao large model.

Specifically, for the Doubao general model pro-32k version, the model inference input price is only 0.0008 yuan/thousand tokens; for the Doubao general model pro-128k version, the model inference input price is 0.005 yuan/thousand tokens.

In other words, users can buy 1.25 million tokens of the Doubao main model for one dollar - about 2 million Chinese characters.

After ByteDance’s Doubao app announced the implementation of the “lowest price in the industry”, Tencent quickly got involved.

On May 17, Tencent Cloud also significantly reduced the price of Hunyuan large models: the inference input prices of the hunyuan-standard model and hunyuan-pro model with a window size of 32K are 0.0069 yuan/thousand tokens and 0.069 yuan/thousand tokens respectively, which is more than 30% lower than the previous standard price.

Although Tencent’s price is a bit high, the important thing is that new users can enjoy a one-year model resource package for free, and the text generation model shares 100,000 tokens.

Whether it is DeepSeek, Zhipu, ByteDance, or Tencent, when announcing price adjustments, they all more or less compare prices using Baidu Wenxin and Ali Tongyi as benchmarks. This actually also indirectly reflects the position of Baidu and Ali in the domestic large model market.

However, the previous few companies never expected that Baidu and Alibaba would react so quickly, and with more favorable prices, they pushed the large model price war to a climax.

On the morning of May 21, Alibaba Cloud announced that the API input price of Qwen-Long, the main GPT-4 model of Tongyi Qianwen, has been reduced from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, a direct reduction of 97%. Users only need to spend 1 yuan to buy 2 million tokens - about 3 million Chinese characters.

However, just a few hours later, Baidu Smart Cloud directly announced that the two main models of Wenxin Model were completely free and effective immediately.

It is reported that Baidu's so-called two main models are the lightweight large models ERNIE Speed ​​and ERNIE Lite launched in March this year, which support 8K and 128K context lengths.

Source: Baidu Smart Cloud Official Account

You downgrade me, I downgrade him, I downgrade him for free. In this big model battle, the players seem to have lost their "sanity".

2. Make the cake bigger and exchange price for land

In fact, price wars are taking place not only in China but also in the overseas large model market.

On May 14, OpenAI officially launched the new generation flagship AI model "GPT-4o" at its spring conference. It not only greatly surpasses GPT-4 Turbo in terms of functionality, but also brings a surprise in terms of price. According to the official introduction, users of the new model GPT-4o do not need to register, and all functions are free.

Before this, free users of ChatGPT could only use GPT-3.5. After the update, users can use GPT-4o for free to perform data analysis, image analysis, Internet search, access app stores, and other operations.

Of course, the free premise is limited to a certain number of messages. Once the specified number is exceeded, the model of free users will be switched back to GPT-3.5. Paid users will receive a higher number of messages, at least 5 times that of free users.

Source: Daily Economic News

As we all know, the cost of training large models is extremely high. With an investment of tens of millions, why would players sell them at a very low price or even for free?

In the industry's view, the current substantial price cuts and free strategies will allow more small and medium-sized enterprises and individual developers to use big model technology. A sufficiently large amount of usage will not only accelerate technological breakthroughs and cost reductions, but also accelerate the penetration of big models in all walks of life and accelerate the commercial development of big models.

In other words, the current price war is essentially a competition among players to compete for customers and expand their market size through lower prices in preparation for future commercialization.

Using price to expand scale is Alibaba Cloud's main strategy to compete for the market this year. After the price cut, Liu Weiguang, senior vice president of Alibaba Cloud Intelligence Group and president of the Public Cloud Division, said: "As China's largest cloud computing company, Alibaba Cloud has significantly reduced the price of large model inference in the hope of accelerating the explosion of AI applications. We expect the number of calls to large model APIs to increase by tens of thousands of times in the future."

When talking about the reason for the price reduction of ByteDance's Doubao large model, Tan Dai, president of Volcano Engine, also said: "Only with large usage can a good model be polished and the unit cost of model reasoning can be greatly reduced." Tan Dai believes that reducing costs is a key factor in promoting large models to fast-forward to the "value creation stage."

Indeed, since the release of ChatGPT at the end of 2022, there have been problems with the commercialization of large models. There are very few companies that really generate a lot of revenue in the field of large models, and most of their revenue depends on financing. Anxiety about commercialization has prompted the occurrence of a price war for large models.

However, price wars not only have obvious benefits, but also obvious disadvantages.

3. Not only the price, but also the technology

In the business world, price wars are nothing new; they are an effective means of attracting more users, increasing user stickiness and market share.

When Taobao and JD.com divided the e-commerce world, Pinduoduo quietly rose with user subsidies and social fission; when Starbucks monopolized the Chinese coffee market, Luckin Coffee also expanded two thousand stores in a year and a half by burning money and gained hundreds of millions of users.

But the money-burning strategy is not always successful. For example, OFO burned through 50 billion yuan and had to exit the market with regret. Another example is LeTV’s complete withdrawal from the Internet competition arena after its capital chain broke.

Today, the price war of large models can promote the implementation of large model applications, allowing more startups to afford large models. But at the same time, price cuts will also cause some small manufacturers who cannot follow the price cuts to gradually exit the market, and large manufacturers with more resources, funds and technical strength will occupy a larger market share.

Of course, the reshuffle effect will not be very obvious in the short term. The real industry reshuffle will begin when the mass market opens up, the performance of large models is gradually recognized by the public, and there are obvious differences in the performance and uses of large models of various companies.

In addition, the reduction will lead to intensified market competition and compressed industry profits, which may also force players to reduce the quality of service while lowering prices, and cause product homogeneity problems.

What’s more important is that free things are often the most expensive.

For example, in the competition stage many years ago, in order to grab user resources, long video platforms basically did not charge fees. But now, the three remaining "iQIYI, Youku and Tencent Video" platforms are not only full of advertisements, but most film and television resources require "membership" to watch.

In the stage of burning money, or before reaching a relative monopoly, this process can be said to provide benefits to society, but the final result is still to achieve a monopoly, and only after the monopoly will the oligarchs really make money. By then, users will have no choice.

Is the current large-scale model market suitable for a price war? Will the future end in the same way as price wars in other fields? It is still unknown, but one thing is certain: the improvement and innovation of technical capabilities are the basis for ensuring the popularization and sustainable development of large-scale model applications.

Only by continuously developing new technologies and new products with independent intellectual property rights that can better meet user needs can players survive longer.

While large-scale model players are competing on price, they should also compete on technology.

Author: Lu Shiming, Editor: Dafeng This article is written by the author of Operation School [Zinc Finance], WeChat public account: [Zinc Finance], original/authorized to be published in Operation School, and any reproduction without permission is prohibited.

The title image is from Unsplash, based on the CC0 protocol.