On May 22, IFLYTEK Announced that iFLYTEK Spark API capability is free. The day before, Alibaba Cloud announced the price reduction of its Tongyi Qianwen GPT-4 main model, while Baidu announced that the two main models of its Wenxin model were free. As the technology competition among big model manufacturers becomes more intense, the industry has obvious signals to start the price war. "Seize the market as soon as possible" is becoming the common strategy of the leading players. By lowering the price threshold, large model manufacturers are expected to attract a broader group of enterprise users, thus further balancing revenue and costs. At the same time, more C-end users are expected to use basic AI applications for free. The huge number of visits will help enterprises further improve their model service capabilities and complete a virtuous circle
◎ Reporter Sun Xiaocheng
The price war of the big model continues to rise in all directions, and new faces continue to join the war.
On May 22, iFLYTEK announced that iFLYTEK's Spark API capability was officially opened for free. Among them, the iFLYTEK Spark Lite API is permanently free of charge, and the price of the iFLYTEK Spark top configuration (Spark3.5 Max) API is as low as 21 yuan/10000 tokens. At present, Baidu Wenxin Yiyan ERNIE-4.0 and Alibaba Tongyi Qianwen Max are priced at 1.2 yuan/10000 tokens.
As for joining the price war, iFLYTEK said that the new price system will help developers reduce the cost of calling, drive product innovation and verification, and address the real world needs.
IFLYTEK's goal: to gather 1 million developers
At the internal annual meeting in February this year, facing 2024, LIU Qingfeng, Chairman of iFLYTEK, put forward the four annual goals of iFLYTEK's big model strategy: leading the base big model, leading the Spark general big model in benchmarking GPT-4 in China; Enabling software and hardware products, enabling 10 million intelligent hardware terminals and more than 100 million independent software users; The industry application is leading, and the market in education, medical care, finance, automobile and other fields ranks first; The developer ecology is leading, bringing together 1 million developers to form the industry's first ecology.
Recently, iFLYTEK has made considerable progress in the field of large models. On April 26, iFLYTEK Spark Model V3.5 "went new", and iFLYTEK officially released the industry's first long text, long text and long voice model. The upgraded iFLYTEK Spark can not only quickly learn massive texts, graphic materials, conference recordings, etc. from various information sources, but also give professional and accurate answers in various industry scenarios.
IFLYTEK has also further upgraded the Spark speech model, with the first multi emotional super anthropomorphic synthesis, which has the ability to express emotions, and introduced the voice reproduction function of one sentence. The Haobo HT 2024 model under GAC Ai'an has taken the lead in carrying iFLYTEK's super anthropomorphic synthetic technology, and was launched globally on April 25.
On April 23, Liu Qingfeng said at the performance briefing that 50% of the R&D investment in 2024 is expected to be used for the construction of large model base; It is expected that in June this year, the iFLYTEK Spark Model will be able to benchmark the current best level of GPT-4turbo. In the future, iFLYTEK's investment around the basic large model may account for 30% of R&D investment, and the proportion of application investment will increase. Although the proportion decreases, the scale of R&D investment around the basic large model will still grow with the growth of the industrial implementation scale.
"With the substantial improvement of the base model capability, the base logic of various AI technology research will be rewritten by the large model technology in the next year or two, and then go to a big step. At the same time, we are also actively studying how to reduce the size of the model through 'distillation' and 'pruning', so as to apply it to mobile phones, homes, cars and other scenes. " Liu Qingfeng said.
Can price reduction bring "multi win"
In the big model price war, before iFLYTEK stepped down, the major Internet giants were in a fierce battle.
On May 21, Alibaba Cloud announced its Qwen Long, the main GPT-4 model of Tongyi Qianwen, and the API input price dropped from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, down 97%. One yuan can buy 2 million tokens, which is equivalent to the text volume of five Xinhua Dictionaries. After the price reduction, the price of Tongyi Qianwen is about 1/400 of the price of GPT-4.
A few hours after Alibaba Cloud's official announcement, Baidu also made a quick move and suddenly announced that the two main models of Wenxin model were free and effective immediately in the afternoon of the same day.
Previously, Byte Beat announced on May 15 that the big model of bean curd was officially opened for external services. The reasoning input prices of the general model of bean curd, pro-32k and pro-128k, in the enterprise market were only 0.0008 yuan/thousand tokens and 0.005 yuan/thousand tokens, respectively, a decrease of 99.3% and 95.8% compared with the industry.
Hualong Securities Research News said that with the further fierce technical competition between large model manufacturers at home and abroad, the signal of opening price war in the large model industry is obvious. By lowering the price threshold, large model manufacturers are expected to attract a broader group of enterprise users, thus further balancing revenue and costs. At the same time, more C-end users are expected to use basic AI applications for free. The huge number of visits will help enterprises further improve their model service capabilities and complete a virtuous circle.
Li Kaifu, CEO of OneThing, said: "To a certain extent, it is expected and should happen that the reasoning cost of the whole industry will be reduced by 10 times every year. Today, it may be at a relatively low point, but I think it is good news to say that the price will be reduced by 10 times every year in the future. Because today's API model calls are still a very low proportion. If they are reduced by 10 times a year, many people can use them. "
However, in Li Kaifu's view, the price of the big model was cut crazily or it was a "lose lose". "I don't think the big model company is so irrational. Because technology is still the most important. If technology is not good, we will never benchmark with such pricing, and we are very proud of our model performance." Li Kaifu said.
When the stock market recovers, open an account first! Intelligent fixed investment, condition sheet, individual stock radar... for you>>
Massive information, accurate interpretation, all in Sina Finance APP
