The big model industry in May was particularly lively. After OpenAI and Google released new developments that shocked the industry last week, the "price war" in the big model circle began, and Byte, Alibaba and Baidu joined the war in a few days. The move of price reduction not only accelerates the application of AI big model inclusive, but also draws people's concern about the industry falling into a strange circle.
The "price war" of the big model is becoming fiercer
Yesterday (21st), Alibaba Cloud announced a message that stirred up ripples in the industry: Qwen Long, the main GPT-4 model of Tongyi Qianwen, dropped the API input price from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, a direct drop of 97%, which means that 2 million tokens can be bought for one yuan, which is equivalent to the text volume of five Xinhua Dictionaries. Alibaba Cloud said that after the price reduction, it is about 400 times of the GPT-4 price, "breaking through the global bottom price".
The so-called token is the smallest unit that can be understood and generated in the big model. It can correspond to a word, a number or a punctuation mark. It is reported that Qwen Long is a long text enhanced version of Tongyi Qianwen. Its performance benchmarking is GPT-4, and the context length is up to 10 million. In addition to the input price dropping to 0.0005 yuan/thousand tokens, the output price of Qwen Long has also dropped by 90% to 0.002 yuan/thousand tokens. In contrast, the input prices of GPT-4, Gemini1.5Pro, Claude3Sonnet and Ernie-4.0 per thousand tokens at home and abroad are 0.22 yuan, 0.025 yuan, 0.022 yuan and 0.12 yuan respectively.
Shortly after Alibaba Cloud announced the price cut, Baidu Smart Cloud announced that the two main models of Baidu Wenxin, ENIRE Speed and ENIRE Lite, are free and effective immediately.
In less than half a day, the major players in the industry unexpectedly threw "heavy bombs" one after another. For a moment, people in the industry sighed that "the volume is crazy" and "it was unexpected that the first 618 was the big model circle".
In fact, the trend of price reduction sweeping large model enterprises is not unique to China. Last week, OpenAI released GPT-4o, which supports free trial. The price of API calls is half that of GPT-4-turbo, which is 5 dollars/million tokens. Since last year, OpenAI has made several rounds of price cuts.
Apply inflection point or acceleration
In fact, it is not just the Internet giant. In the past two weeks, star entrepreneurial teams such as OneThing and Zhipu AI have also been involved in this price war. A week ago, a new price system was launched on the open platform of the smart spectrum big model. The call price of the entry-level product GLM-3 Turbo model was reduced by 80%, and 1 million tokens could be purchased for 1 yuan. On the same day, Kai Fu Lee, the founder and CEO of OneThing, announced that the training cost of OneThing's 100 billion parameter model had been optimized, and the year-on-year decrease had doubled.
"This is good news for the whole industry". Yesterday, in response to the price reduction action of friends, Kai Fu Lee said, "The industry can expect to reduce the reasoning cost by 10 times every year, and it is inevitable." He also predicted that the reduction of the reasoning cost of the big model will push China's AI big model into the stage of landing as the king, This year will usher in the "first year of big model application explosion".
Liu Weiguang, senior vice president of Alibaba Cloud Intelligent Group and president of Public Cloud Business Department, also explained that this round of price reduction benefited from the dividend and scale effect after the scale of public cloud technology, bringing cost and performance advantages.
However, the cost reduction brought by technological progress is an important reason for this round of price reduction, but it is by no means the only driver. Tan Dai, president of Volcano Engine, said when talking about the reason why the byte model took the lead in provoking a "price war", on the one hand, the price could be reduced for technical reasons, and on the other hand, because the market was crying out for price reduction, the platform must make trial and error costs very low to attract everyone to use. Tan Dai especially mentioned that the expansion of user scale will also "grind" a good model, further significantly reducing the unit cost of model reasoning.
Irrationally burning money or "losing both"
Big companies have "cut meat", which has brought considerable cost pressure to start-ups. Mr. Yuan, the founder of an AI startup, said helplessly: "The big model APIs are free. What's the point of doing performance optimization? If it is meaningful, it is to reduce the subsidy loss."
Li Di, a Xiaobing company, said that if the big model could not enable AI companies to obtain benign economic benefits from it, they could only rely on free services to gain customers and expand the market scale, instead of winning the market through innovation, then such a business model would only be a flash in the pan. "It is only a temporary measure to compete for the bottom down, and the big model that can be priced up will have a future," said another senior practitioner in the industry.
Li Kaifu also said that although price reduction is the trend of the times, he still believes that the irrational price war is a "lose lose" approach. When his team believes that their technology is worth it, they will adhere to a suitable price, and will never rely on subsidies or losses to do business.
The big model industry in May was particularly lively. After OpenAI and Google released new developments that shocked the industry last week, the "price war" in the big model circle began, and Byte, Alibaba and Baidu joined the war in a few days. The move of price reduction not only accelerates the application of AI big model inclusive, but also draws people's concern about the industry falling into a strange circle.
The "price war" of the big model is becoming fiercer
Yesterday (21st), Alibaba Cloud announced a message that stirred up ripples in the industry: Qwen Long, the main GPT-4 model of Tongyi Qianwen, dropped the API input price from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, a direct drop of 97%, which means that 2 million tokens can be bought for one yuan, which is equivalent to the text volume of five Xinhua Dictionaries. Alibaba Cloud said that after the price reduction, it is about 400 times of the GPT-4 price, "breaking through the global bottom price".
The so-called token is the smallest unit that can be understood and generated in the big model. It can correspond to a word, a number or a punctuation mark. It is reported that Qwen Long is a long text enhanced version of Tongyi Qianwen. Its performance benchmarking is GPT-4, and the context length is up to 10 million. In addition to the input price dropping to 0.0005 yuan/thousand tokens, the output price of Qwen Long has also dropped by 90% to 0.002 yuan/thousand tokens. In contrast, the input prices of GPT-4, Gemini1.5Pro, Claude3Sonnet and Ernie-4.0 per thousand tokens at home and abroad are 0.22 yuan, 0.025 yuan, 0.022 yuan and 0.12 yuan respectively.
Shortly after Alibaba Cloud announced the price cut, Baidu Smart Cloud announced that the two main models of Baidu Wenxin, ENIRE Speed and ENIRE Lite, are free and effective immediately.
In less than half a day, the major players in the industry unexpectedly threw "heavy bombs" one after another. For a moment, people in the industry sighed that "the volume is crazy" and "it was unexpected that the first 618 was the big model circle".
In fact, the trend of price reduction sweeping large model enterprises is not unique to China. Last week, OpenAI released GPT-4o, which supports free trial. The price of API calls is half that of GPT-4-turbo, which is 5 dollars/million tokens. Since last year, OpenAI has made several rounds of price cuts.
Apply inflection point or acceleration
In fact, it is not just the Internet giant. In the past two weeks, star entrepreneurial teams such as OneThing and Zhipu AI have also been involved in this price war. A week ago, a new price system was launched on the open platform of the smart spectrum big model. The call price of the entry-level product GLM-3 Turbo model was reduced by 80%, and 1 million tokens could be purchased for 1 yuan. On the same day, Kai Fu Lee, the founder and CEO of OneThing, announced that the training cost of OneThing's 100 billion parameter model had been optimized, and the year-on-year decrease had doubled.
"This is good news for the whole industry". Yesterday, in response to the price reduction action of friends, Kai Fu Lee said, "The industry can expect to reduce the reasoning cost by 10 times every year, and it is inevitable." He also predicted that the reduction of the reasoning cost of the big model will push China's AI big model into the stage of landing as the king, This year will usher in the "first year of big model application explosion".
Liu Weiguang, senior vice president of Alibaba Cloud Intelligent Group and president of Public Cloud Business Department, also explained that this round of price reduction benefited from the dividend and scale effect after the scale of public cloud technology, bringing cost and performance advantages.
However, the cost reduction brought by technological progress is an important reason for this round of price reduction, but it is by no means the only driver. Tan Dai, president of Volcano Engine, said when talking about the reason why the byte model took the lead in provoking a "price war", on the one hand, the price could be reduced for technical reasons, and on the other hand, because the market was crying out for price reduction, the platform must make trial and error costs very low to attract everyone to use. Tan Dai especially mentioned that the expansion of user scale will also "grind" a good model, further significantly reducing the unit cost of model reasoning.
Irrationally burning money or "losing both"
Big companies have "cut meat", which has brought considerable cost pressure to start-ups. Mr. Yuan, the founder of an AI startup, said helplessly: "The big model APIs are free. What's the point of doing performance optimization? If it is meaningful, it is to reduce the subsidy loss."
Li Di, a Xiaobing company, said that if the big model could not enable AI companies to obtain benign economic benefits from it, they could only rely on free services to gain customers and expand the market scale, instead of winning the market through innovation, then such a business model would only be a flash in the pan. "It is only a temporary measure to compete for the bottom down, and the big model that can be priced up will have a future," said another senior practitioner in the industry.
Li Kaifu also said that although price reduction is the trend of the times, he still believes that the irrational price war is a "lose lose" approach. When his team believes that their technology is worth it, they will adhere to a suitable price, and will never rely on subsidies or losses to do business.
