It is better to buy well than to sell well. There are still a series of problems worth discussing behind the collective price reduction of large model manufacturers.
On May 15, the volcano engine took the lead and announced that the main model of bean curd was priced at 0.0008 yuan/thousand tokens in the enterprise market, which was 99.3% cheaper than the industry. Its practice of precisely sniping peers directly set off a low price hand to hand battle for large model manufacturers. Ali, Baidu, iFLYTEK and Tencent have stepped forward to fight. On May 21, Alibaba Cloud officially announced that the input price of Qwen Long, the main model of Tongyi Qianwen, dropped to 0.0005 yuan/1000 tokens, down 97%; Only a few hours later, Baidu Intelligent Cloud presented the killer technique, announcing that the two main models of Wenxin big model, ERNIE Speed and ERNIE Lite, were free. Since Baidu, the big model has been completely linked to free. On May 22, iFLYTEK announced that the iFLYTEK Spark Lite API will be open for free forever. In the afternoon, Tencent Cloud announced a new big model upgrade plan, and the hybrid lite model, one of the main models, was adjusted to be free of charge. In just one week, the big model has moved from the "Li" era to the "free" era. The underlying driving force behind the apparent price reduction is technology. After more than a year of technology catch-up, domestic large model manufacturers have achieved breakthroughs in computing power, reasoning, algorithms and other aspects, thus achieving cost reduction in technology. In addition, the large-scale advantages brought by cloud computing of large factories jointly triggered a wave of price cuts. On the other hand, it also confirms that the big model has entered a new stage of availability from the release demo. Tan Dai, president of Volcano Engine, mentioned a standard when talking about the release and price reduction time of Doubao large model: "model capability is ready". At present, the premise for the large-scale open use of major model manufacturers is that the model capability has passed the test and can be stably supplied. On closer examination, the low price and free price of the big model manufacturers are more like the cheese that lures mice out of their holes. This kind of free product has many restrictions. The products with the highest price reduction in Alibaba and Baidu are both lightweight model versions, which are only suitable for short-term use by SMEs and developers with low frequency of use, less reasoning and less complex task processing. In this case, the "Internet" means of low price and free become the customer acquisition strategy of large model manufacturers. While obtaining more data to optimize the model effect, they try to convert to a higher paid version through tasting new products. It is better to buy well than to sell well. There are still a series of problems worth discussing behind the collective price reduction of large model manufacturers. Sell AI big model with Internet free method From the perspective of users, there may be two types of potential beneficiaries of large model price reduction: developers and enterprises. Although the large-scale price reduction in the industry is the first time, as early as last year, major manufacturers attracted AI entrepreneurs and teams to participate by winning the Hacksong contest and sending tokens. At that time, a frequent visitor of Hacksong told Photon Planet, "To participate in a contest is to collect wool, and to take a token is to take nothing". Collecting wool can really reduce the cost of starting a business. Lowering the price is friendly to developers, especially independent developers. This may mean that developers can run more tests and get more feedback data, thus shortening the product launch cycle and further improving 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 was a polarized voice among developers and enterprises. One side agrees with the price reduction of domestic large models and believes that developers and enterprises can continue to collect wool. After all, there are many cases of shell application products in the market now; On the other hand, the price reduction of large model manufacturers is lack of sincerity, and the large price reduction is only for small-scale models. Although they claim to be able to benchmark GPT-4, in fact, GPT-3.5 is not as good as GPT-3.5, and the model level is not up to standard, so it is impossible to operate in the actual production environment. The apparent price reduction of big model manufacturers is actually a mystery behind it. This is like giving you a limited time experience card of cloud disk. Just three seconds after watching the high-definition video, you will be prompted to upgrade your VIP. It also happens that just five seconds after experiencing the fast download, you will be prompted to upgrade your membership permissions. The taste of the big model is also very good. It attracts developers and enterprises to use it with the gimmick of price reduction and free. As soon as they get started, they are started with key indicators such as card call speed, reasoning speed, task processing capacity, etc. Moreover, Photon Planet further found that the price reduction strategy of large model manufacturers did not have a substantial impact on commercialization. The result is that the price of the big model manufacturer has dropped and the money has not been lost. Insiders of a large factory told Photon Planet that the main way to commercialize large models is to take orders from to B. Similar to the cooperation mode of SaaS and cloud, there are two ways: case by case and cooperation commission. Among them, case by case is a more mainstream way of cooperation, that is, existing customers of a large model manufacturer will begin to try the large model of the manufacturer because they are already using the cloud and SaaS products of the manufacturer. Accordingly, 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 situations: the big model has become a value-added element of SaaS products or project cooperation. The big model itself does not pay, but in order to hedge costs, the big model manufacturer has to in turn raise the price of SaaS and project cooperation. The wool finally came out of the sheep. As the price went up and down, the big factory did not lose, but earned. The price of the big model has been reduced, and then what? Perhaps the impact of the domestic price war for big models is that from now on, big models are officially equal to "free". This will become a watershed. In the past two years, the AI native product logic of "online charging" that many entrepreneurs and teams tried to establish has been challenged again. Around the corner, the business logic of the Internet once again dominated the development of the big model. No matter at home or abroad, there has always been a state of model mixing in the industry. Essentially, each major 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 use scenarios. A similar situation also occurred in China. We learned that in the process of developing WPS AI functions, Kingsoft Office took turns to try MiniMax, Zhipu AI, Wenxin Yiyan, Shangtang Ririxin, Tongyi Qianwen and other big model capabilities, and built its own platform by understanding the advantages of each big 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, test the ability of different models, and optimize the ability of large models in different tasks. This not only tests the cost, but also avoids excessive dependence on a single product. So far, large model products are often criticized for low user stickiness. Compared with subscription fees, it is difficult to retain customers in the way of API call fees. The same is true for the case by case charging mode on the enterprise side. The enterprise's use of a manufacturer's large model cycle depends on the order cycle. Customers follow the orders. They can use bytes today and Alibaba tomorrow. The essence of price reduction is to accelerate the implementation of large models. The big model should not only write poems and paintings, but also "go to the grassroots". Behind the price reduction is the cooperation case of reaching thousands of industries and obtaining a larger sample size, from which common characteristics are extracted to form a reasonable and efficient industrial standard of large model. When the big model manufacturers are back on the same starting line again, under the condition that the capability level of each model is similar and the price is comparable, the common task they have to face is how to retain customers. From the perspective of large model customers, they prefer to reduce their dependence on a single model through hedging. Under such psychological drive, the future large-scale model model can refer to the procurement methods of SaaS and cloud products. A company can purchase multiple large-scale model companies' products internally, and different product lines and business departments may also use different large-scale models. If you win the price, you win everything? In retrospect, the big model has gone from hundred models, parameters and long texts to the current price. Past experience tells us that price cannot be the only determinant. The prices given by the big model manufacturers are not very competitive in the market, even if we don't talk about the right version of what enterprises and developers get. The open source model is more cost-effective than the domestic model. A domestic staff member in charge of e-commerce agency operations told Photon Planet that up to now, his business department has purchased AI related pay products such as ChatGPT and Midjournal, and now the bottom layer uses the open source and commercially available Llama 3. The reason why some companies and developers prefer to deploy open source models is that, on the one hand, the ability of foreign open source models such as Llama has been catching up with the level of the strongest version of ChatGPT, and some common scenario capabilities are sufficient for business. On the other hand, it is more flexible to deploy and fine tune the model from the beginning for later business adjustment. In addition, the discovery of Photon Planet has also derived the role of middleman between the closed source large model manufacturers and the open source community. A puzzling phenomenon is spreading in the big model industry: the API price sold by the big model distributor is cheaper than the original price. Taking the Deepbricks platform abroad as an example, the official input price of OpenAI for the latest GPT-4o model is $5/1M tokens, while the price of Deepbricks itself is only $2/1M tokens. If these middlemen can really update the model in real time and achieve low prices, they may attract a group of developers and enterprises to use them in the future. (Source: Deepbricks official website) Jia Yangqing, founder of Lepton AI and former vice president of Alibaba, believes that enterprises are not cost driven when using AI. 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 is wasteful to use them cheaply. If the simple price is not attractive, which big model will be used by the customer depends on what? An entrepreneur of middleware said to Photon Planet: "The most important thing is the model effect. If the model effect is too poor, it can't be used even if it is too cheap." And overseas AI entrepreneurs directly told Photon Planet that ChatGPT is used abroad because of its strong ability; Wenxin is used in China because it can meet the requirements of compliance. Therefore, price is only one of the factors for enterprises to choose the big model. Similarly, in the era of cloud computing and SaaS, it is not the low price that can keep customers, but the deeper binding relationship or interest relationship. For example, when the enterprise adopts the Doubao model of volcanic engine, whether it can enjoy the preferential right in dithering investment; Access to Tongyi Qianwen, can its products connect with Alibaba Ecology and get more resource support? When enterprise users choose large models, they are also weighing the respective advantages of manufacturers. The ability of the big model is the second. More important is how much growth can be brought to its business by choosing this manufacturer, and how much revenue can be obtained under the manufacturer's industrial chain. In the end, we still need to talk about the results. As Jia Yangqing said, "Maybe it is not the cheapest way to win the business war, but the way to win profits."
