Big model price cuts, a war with no winners

In just one week, large models have entered the "free" era from the "centimeter" era. Behind the collective price cuts by large model manufacturers, there are still a series of issues worth discussing.

The price trend of 6.18 rolls has reached the large model circle.

On May 15, Volcano Engine took the lead and announced that the price of Doubao's main model in the enterprise market was 0.0008 yuan/thousand tokens, which was 99.3% cheaper than the industry average. Its precise attack on its peers directly triggered a price war among large model manufacturers.

Alibaba, Baidu, iFLYTEK and Tencent have all stepped forward to respond to the challenge.

On May 21, Alibaba Cloud officially announced that the input price of Qwen-Long, the main model of Tongyi Qianwen, was reduced to 0.0005 yuan/thousand tokens, a direct drop of 97%; only a few hours later, Baidu Smart Cloud launched its ultimate move and announced that the two main models of Wenxin Model, ERNIE Speed ​​and ERNIE Lite, were completely free.

Since Baidu, big models have been completely associated with free.

On May 22, iFLYTEK announced that the iFLYTEK Spark Lite API would be permanently open for free. In the afternoon, Tencent Cloud announced a new large model upgrade plan, and the Hunyuan-lite model, one of the main models, was adjusted to be completely free.

In just one week, the big model went from the "centi" era to the "free" era.

On the surface, it is a price reduction, but the fundamental driving force behind it comes from technology. After more than a year of technological catch-up, domestic large model manufacturers have achieved breakthroughs in computing power, reasoning, algorithms and other aspects, thereby achieving cost reduction in technology. Coupled with the scale advantages brought by cloud computing of large manufacturers, it has jointly triggered the price reduction trend.

On the other hand, it also indirectly confirms that the big model has entered a new stage of availability from the demo at the press conference. Tan Dai, president of Volcano Engine, mentioned a standard when talking about the release and price reduction of Doubao big model: "The model capabilities are ready." At present, the premise for major model manufacturers to open up for large-scale use is that the model capabilities have passed the test and can be stably supplied.

Upon closer inspection, the low prices or free services offered by large model manufacturers are more like cheese to lure mice out of their holes.

This free service comes with many restrictions. The products with the largest price cuts by Alibaba and Baidu are their lightweight model versions, which are only suitable for short-term use by small and medium-sized enterprises and developers with low usage frequency, low inference volume, and low complex task processing volume.

In this case, "Internet" means such as low prices and free services have become customer acquisition strategies for large model manufacturers. On the one hand, they obtain more data to optimize model effects, and on the other hand, they try to convert to higher-end paid versions by trying out new products.

It is better to sell well than to buy well. There are still a series of issues worth discussing behind the collective price reduction of large model manufacturers.

1. Selling AI models for free on the Internet

From the user's perspective, there may be two types of potential beneficiaries of the price reduction of large models: developers and enterprises.

Although this is the first time that the industry has seen a large-scale price cut, as early as last year, major companies have been attracting the participation of AI entrepreneurs and teams by offering tokens to those who win hackathon competitions.

At that time, a hackathon regular told Photon Planet, "Participating in the competition is just a way to get free stuff, it would be a waste if you don't get the tokens."

Taking advantage of the market can indeed reduce the cost of starting a business. Lowering prices is friendly to developers, especially independent developers. This may mean that developers can run more rounds of tests and obtain more rounds of feedback data, thereby shortening the product launch cycle and further increasing the possibility of entrepreneurial success.

But the premise is to meet the needs of developers and enterprises. Photon Planet learned that after the news of the price reduction came out, there were polarized voices among developers and enterprises.

One side agrees with the price reduction of large domestic models, and believes that developers and enterprises can continue to take advantage of them. After all, there are many cases of shell application products on the market. The other side feels that the price reduction of large model manufacturers lacks sincerity, and the large price reductions are for small-scale models. Although they claim that their level can match GPT-4, in fact, they are not even as good as GPT-3.5. The model level is not up to standard and cannot be run in an actual production environment.

The apparent price cuts by large model manufacturers are actually a ulterior motive. It is like giving you a limited-time cloud disk experience card, and you are reminded to upgrade to VIP after watching a three-second high-definition video. It is also like you are reminded to upgrade your membership privileges after experiencing a five-second ultra-fast download.

The trial of large models is similar. They use the gimmicks of price reduction and free of charge to attract developers and enterprises to use them. However, as soon as they get started, they start to encounter key indicators such as call speed, inference speed, and task processing volume.

Moreover, Photon Planet further discovered that the price reduction strategy of large model manufacturers did not have a substantial impact on commercialization. The result presented is that large model manufacturers have reduced prices and made a lot of money.

An insider of a large company told Photon Planet that the main commercialization method of large models is to get B2B orders. Similar to the cooperation model of SaaS and cloud, there are two methods: case by case and cooperation commission.

Among them, case by case is the more mainstream cooperation mode, that is, the existing customers of large model manufacturers will start to try the large models of a certain manufacturer because they are already using the cloud and SaaS products of that manufacturer. Correspondingly, in order to retain customers, large model manufacturers will also add AI functions to their own SaaS and cloud products.

This may lead to the following situation: Big models become a value-added factor of SaaS products or project cooperation. Big models themselves are free, but in order to hedge costs, big model manufacturers have to increase the price of SaaS and project cooperation. The wool is ultimately paid by the sheep, and the price rises and falls, and the big manufacturers not only do not lose money but also make money.

2. The price of large models has dropped, so what?

Perhaps the impact of the domestic large-model price war is that from now on, large models are officially equated with "free".

This will be a watershed moment. In the past two years, the logic of AI native products that entrepreneurs and teams have tried to establish, which is to charge as soon as they go online, has been challenged again. After many twists and turns, the business logic of the Internet has once again dominated the development of large models.

Whether at home or abroad, there has always been a state of mixed use of models in the industry. The essence is that each large model has its own strengths. For example, ChatGPT is good at reasoning, and Claude is good at writing. It is based on the characteristics of different models that users will call corresponding models in different usage scenarios.

A similar situation also occurred in China. We learned that in the process of developing the WPS AI function, Kingsoft Office tried MiniMax, Zhipu AI, Wenxin Yiyan, SenseTime Daily, Tongyi Qianwen and other large model capabilities in turn, and built its own platform by understanding the advantages of each large model.

Last year, a domestic data governance company told Photon Planet that they would also run a large number of models in the early stage to test the capabilities of different models and select the best capabilities of large models in different tasks. This not only conducts cost testing, but also avoids over-reliance on a single product.

Until now, large-scale model products have often been criticized for their low user stickiness. Compared with subscription charges, charging by API calls is inherently difficult to retain customers.

The same is true for the case-by-case charging model on the enterprise side. The period for an enterprise to use a certain manufacturer's large model depends on the order cycle. Customers follow the orders and can use ByteDance today and Alibaba tomorrow.

The essence of price reduction is to accelerate the implementation of large models. Large models cannot just stay in writing poems and paintings, but also need to "go down to the grassroots." Behind the price reduction is to reach thousands of industries and obtain cooperation cases with larger sample capacity, extract common features from them, and form reasonable and efficient large model industry standards.

When the big model manufacturers were brought back to the same starting line, with their models having similar capabilities and prices, the common issue they faced became how to retain customers.

From the perspective of big model customers, they hope to reduce their reliance on a single model through hedging. Driven by this mentality, the future big model model can refer to the procurement method of SaaS and cloud products. A company can purchase products from multiple big model companies, and different product lines and business departments may also use big models from different companies.

3. If you win on price, do you win everything?

Looking back at history, large models have gone from hundreds of models, parameters, and long texts to the current price. Past experience tells us that price cannot be the only determining factor.

Even without considering whether what enterprises and developers get is correct or not, the prices offered by large model manufacturers are not very competitive in the market.

Open source large models are more cost-effective than domestic large models. A domestic staff member in charge of e-commerce agency operations told Photon Planet that so far, his business department has purchased paid AI-related products such as ChatGPT and Midjourney, and now uses the open source and commercially available Llama 3 at the bottom.

Some companies and developers prefer to deploy open source models because, on the one hand, the capabilities of foreign open source models such as Llama have been catching up with the level of the strongest version ChatGPT, and some general scenario capabilities are sufficient for business. On the other hand, deploying and fine-tuning models from scratch is more flexible for later business adjustments.

In addition, Photon Planet found that there are middlemen between the closed-source big model manufacturers and the open source community. A puzzling phenomenon is spreading in the big model industry: the API prices sold by big model distributors are cheaper than the original manufacturers.

Taking the overseas Deepbricks platform as an example, the latest GPT-4o model, the official OpenAI input price is 5 US dollars/1M tokens, while Deepbricks' own selling price is only 2 US dollars/1M tokens. If these middlemen can really achieve the ability to update models in real time and keep the price low, they may attract a group of developers and enterprises to use them in the future.

(Image source: Deepbricks official website)

Jia Yangqing, founder of Lepton AI and former vice president of Alibaba, believes that when enterprises use AI, it is not cost-driven. It is not because APIs are expensive that no one uses them, but because enterprises must first figure out how to use them to generate business value. Otherwise, it will be a waste no matter how cheap it is.

If price alone is not attractive, what will determine which large model a customer uses?

A middleware entrepreneur told Photon Planet: "The most important thing is to look at the model effect. If the model effect is too poor, it cannot be used no matter how cheap it is."

There are also overseas AI entrepreneurs who directly told Photon Planet that ChatGPT is used abroad because of its strong capabilities, and Wenxin Yiyan is used in China because it can meet compliance requirements.

Therefore, price is only one factor in companies choosing large models.

Similarly, in the era of cloud computing and SaaS, it is often not low prices that can retain customers, but deeper binding relationships or interest relationships. For example, when an enterprise adopts the Doubao model of Volcano Engine, can it enjoy preferential rights in Douyin traffic? If it connects to Tongyi Qianwen, can its products be connected to the Alibaba ecosystem and obtain more resource support?

When enterprise users choose big models, they are also weighing the advantages of each manufacturer. The capabilities of the big models are secondary. What is more important is how much growth this manufacturer can bring to their business and how much profit they can get under the manufacturer's industrial chain.

In the end, the results will speak for themselves. As Jia Yangqing said, "Maybe it's not the cheapest way to win a business war, but the way to win a profit that can be implemented."