Chinese startup DeepSeek has surprised the AI industry with a real breakthrough thanks to its powerful and extremely affordable AI models. The company's two key language models – DeepSeek R1 and DeepSeek V3 – have the potential to not only compete with market leaders such as OpenAI, Google, Anthropic, xAI, but also surpass them, while using significantly fewer resources. One of the biggest shocks to the industry was the astonishingly low prices for these models.
What else is known
Traditional technology vendors such as OpenAI, Anthropic, and Google offer different versions of LLM, which vary in price and performance. For example, OpenAI sells a low-cost version of GPT-4o called the GPT-4o Mini. DeepSeek, in turn, takes a similar approach, where V3 is the cheaper model, and users have to pay more for R1.
Model prices are based on the amount of input and output that users enter and receive. The unit of measurement is — million tokens, where tokens are words or combinations of characters and punctuation marks that the LLM generates when generating text.
The input for V3 is ten times cheaper than for GPT-4o Mini, and twenty times cheaper than for Anthropic's Claude 3 Haiku. The price of the V3 results is also lower than that of the GPT-4o Mini. At the same time, the V3 raw data is slightly more expensive, but still remains a bargain compared to models from Google and Amazon.
DeepSeek has even more attractive prices for the R1 model, which is the cheapest among the most powerful AI models. For example, the Gemini 1.5 Pro, although slightly cheaper, costs significantly more than the R1. Even Amazon’s Nova Pro, the cheapest top model in the US, cannot match its Chinese competitor.
Given these figures, it is not surprising that DeepSeek has caused a great stir in the American technology industry. DeepSeek models can not only compete on price, but also demonstrate a high level of performance, making it difficult to justify the high prices of other vendors.