DeepSeek permanently reduces the price of its flagship V4 Pro model by 75%

The startup DeepSeek wants to become the ‘cost-effective’ option for artificial intelligence (AI) agents and to do so it has introduced a discount of 75 percent that will become permanent starting next week after the end of the discount period this May 31st.

From the DeepSeek website you can read that deepseek-v4-pro model API price will be “officially adjusted” at a quarter of the original price after the discount promotion ends on May 31 of this year.

The V4 Pro model now offers a price range from 0.003625 to 0.87 dollars per million tokens (the smallest unit of text that the model recognizes as a word, number or punctuation mark), compared to the previous range of $0.0145 to $3.48 per million tokens.

The decision made by the company comes after the launch of its flagship models V4 Pro and V4 Flash just a month ago as “the era of low-cost 1 million context length”and which directly hits AI models from American laboratories in price, such as OpenAI’s ChatGPT or Google’s Gemini 3.5 Flash launched last week, as a more affordable alternative for enterprise accounts or advanced users who consume millions of tokens a day.

The difference in costs that DeepSeek has opened compared to OpenAI, Anthropic and Google translates into the fact that DeepSeek V4 Pro is approximately 34 times cheaper than OpenAI’s GPT-5.5 ($0.87 versus $30 per million tokens), while compared to Google’s Gemini 2.5 Pro, it is up to 13 times cheaper ($11.37 per million tokens).

A price reduction display that shows that Software optimization can be as important or more important than brute force of some AI models like ChatGPT, Akamai CTO Robert Blumofe maintains in an article last year, in which he states that “there is obviously limited advantage in the hardware-based scaling approach. The language model space is entering an era of optimization.”

According to the technical report published by KuCoin on DeepSeek V4 Pro, the feasibility for the cost reduction is due to the fact that it “reduces the computational requirements for inference to only 27 percent of the previous generation and cuts memory usage by a tenth, thanks to architectural innovations like a hybrid attention mechanism“.