"Doubao" lowers the price line, the global large model opening price-performance ratio
Written by: Mu Mu
The big models are also starting to engage in price wars.
On May 15, Volcano Engine, a subsidiary of ByteDance, released the Doubao Big Model. In addition to the Doubao APP for C-end users, which can use the model for free, the Doubao Big Model will bring the B-end price to the lowest in the industry.
According to Tan Dai, president of Volcano Engine, the price of Doubao's main model (≤32K) in the enterprise market is only 0.0008 yuan/thousand tokens. 0.8 cents can process more than 1,500 Chinese characters, which is 99.3% cheaper than the industry average.
Before Doubao debuted with high cost performance, Tongyi Qianwen and Zhipu AI, DeepSeek and many other domestic large models have begun to "roll up" prices, and the Hundred Model War has entered a new stage with collective price cuts. As Tan Dai said, reducing costs is a key factor in pushing large models to the "value creation stage".
"Doubao" brings B-end price to a new low in the industry
The predecessor of Doubao Big Model is Yunlark Big Model, which is also the first big model based on Transformer architecture released by ByteDance in August 2023. Half a year later, Doubao Big Model not only released a whole family bucket, but also reduced the price for B-end users in the industry.
The price of Doubao main model in the enterprise market is only 0.0008 yuan/1,000 tokens, and 0.8 cents can process more than 1,500 Chinese characters, which is 99.3% cheaper than the industry price. Based on this calculation, 1 yuan can buy 1.25 million tokens of Doubao main model, which is about 2 million Chinese characters, equivalent to three volumes of "Romance of the Three Kingdoms". The 128K Doubao general model only costs 0.005 yuan/1,000 tokens, which is 95.8% cheaper than the industry price.
You should know that the input of GPT-4 Turbo is 0.01 USD for 1000 Tokens and the output of 1000 Tokens is 0.21 RMB. In comparison, ByteDance directly "cuts the price" and can be called the Pinduoduo of AI.
Not only "Doubao", many large models in China are seeing price cuts.
Not long ago, Baidu released the lightweight version of Wenxin Big Model, in which the price of ERNIE Tiny version dropped to 0.001 yuan per thousand tokens, equivalent to 1 yuan for 1 million tokens.
In May this year, Zhipu AI's large model commercial price also dropped significantly. The entry-level product GLM-3 Turbo model call price was reduced by 80%, from 5 yuan/million tokens to 1 yuan/million tokens, which is enough to allow more companies and individuals to use this entry-level product.
The price of Zhipu AI's large model
May Xiaobai NavigationOn the 6th, DeepSeek, an AI company under the well-known domestic private equity giant Huanfang Quantitative, released the new second-generation MoE large model DeepSeek-V2. The DeepSeek-V2 API is priced at 1 yuan per million tokens input and 2 yuan per output (32K context).
On May 9, Alibaba Cloud officially released Tongyi Qianwen 2.5. According to the evaluation results of OpenCompass, Tongyi Qianwen 2.5 scored on par with GPT-4 Turbo. At the same time, individual users can use it for free through the App, official website and mini program.
On May 14, Tencent's Hunyuan Wenshengtu large model was directly open sourced and free for commercial use.
Overseas, OpenAI's newly released GPT-4o has also been significantly reduced in price. Not only is it free for all users, but the API call price is also half that of GPT-4-turbo released in November last year, but the speed is twice as fast. This is the third price reduction for OpenAI's large model products.
The input and output prices of Mistral Large, a large model of French artificial intelligence company Mistral AI, are currently about 20% cheaper than GPT-4 Turbo, which once attracted widespread attention.
Whether at home or abroad, large models are collectively experiencing price cuts.
Large model reduces costs and increases efficiency through application
The "price war" among manufacturers has already begun. Half a year ago, people knew that large model training was very expensive. Why was it that in just half a year, manufacturers were able to "drive down" prices and raise prices?
Tan Dai, president of Volcano Engine, believes that reducing costs is a key factor in pushing large models to the "value creation stage". For small and medium-sized enterprise customers, an important consideration for calling large models is cost. Tan Dai revealed that ByteDance has many optimization methods in various technical aspects such as model structure, training, and production to achieve price reduction.
OpenAI CEO Sam Altman is also proud that people can use ChatGPT without seeing ads. "One of our key missions is to provide AI products to people for free."
Indeed, low prices are helping large model R&D companies seize market opportunities and gain a foothold. The increase in user volume can in turn help R&D companies train better models. So, has the cost of training large models really decreased?
When GPT-4 was released last year, Sam Altman revealed that the training cost of OpenAI's largest model was "well over $50 million." According to the 2024 Artificial Intelligence Index Report released by Stanford University, it is estimated that OpenAI's GPT-4 training cost is $78 million.
The high cost of training large models also directly pushes up the usage fees, directly blocking many corporate users.
However, researchers are looking for lower-cost training methods. Last year, researchers from the National University of Singapore and Tsinghua University proposed a framework called VPGTrans to train high-performance multimodal large models at a very low cost. Compared with training the visual module from scratch, the VPGTrans framework can reduce the training cost of BLIP-2 FlanT5-XXL from 19,000+ RMB to less than 1,000 RMB.
In the domestic large-scale models, R&D personnel have also found ways to reduce costs and increase efficiency in various aspects. After DeepSeek-V2 improved the quality of the data set and optimized the architecture, the AI heterogeneous computing platform "Bai Ge" increased the throughput of training and reasoning scenarios by up to 30% and 60%.
In addition to the training process, some infrastructure for large model training - chips are also falling in price. For example, the price reduction of Nvidia's AI chip Nvidia A100 directly reduced the cost of large model training by about 60%.
The most direct impact of the big model price war is that the application landing has begun to accelerate. On the Doubao platform, more than 8 million intelligent agents have been created. More than 3 million apps based on the GPT model have been created on the GPT Store.
In just half a year, the era of spending money to compete on large model performance seems to have become a thing of the past. Nowadays, as the prices of large models of various manufacturers drop, market users pay more attention to which large model is more affordable and easy to use. This will promote the faster implementation of large model applications in scenarios and businesses.
The article comes from the Internet:"Doubao" lowers the price line, the global large model opening price-performance ratio
The old public chain Fantom is rising again. Written by: Daniel LicryptocurrencyIn the market, the old altcoin FTM has once again attracted the attention of investors. What is particularly striking is that in March alone, the price of FTM has soared by 150%. What is even more surprising is that even in the recent week, when Bitcoin and Ethereum were affected by Japan’s…
相关文章
