Among Well-Known Models, Deepseek’s New AI Model Is by Far the Least Expensive to Operate

According to a research firm, a version of the flagship AI model from the Chinese startup DeepSeek is more than 100 times less expensive to run than Anthropic’s Claude Fable 5 and by far the least expensive to run on benchmark tests among well-known models worldwide.

DeepSeek, which is reportedly getting ready for a possible initial public offering (IPO), formally unveiled its V4-Flash model on Friday. This is the company’s most recent attempt to recover traction by providing ultra-low-cost AI options.

Early in 2025, the startup’s R1 model gained widespread attention, which led to a selloff in international technology stocks and raised concerns about the substantial sums of money American businesses were investing in artificial intelligence.

Research firm Artificial Analysis claims that DeepSeek’s V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens. A token is a data unit used to quantify the use of AI.

The average cost of V4-Flash was calculated by San Francisco-based Artificial Analysis to be 3 cents per test, while Kimi K3 from Chinese competitor Moonshot AI costs 86 cents, GPT-5.6 Sol from OpenAI costs $1.86, and Claude Fable 5 costs $3.15.

Because the comparison takes into consideration the quantity of data a model needs process and produce in order to do a task, it offers a more accurate assessment of value than pricing alone. Even if a model has a cheap headline price, it may still be costly if it takes a lot more steps to get an answer.

Before being swiftly overtaken by numerous local competitors, including tech behemoths like ByteDance and Alibaba (9988.HK), as well as other startups like Moonshot, MiniMax, and Z.AI, DeepSeek dominated most of the news on Chinese AI development. All are competing with American tech companies for widespread adoption, focusing on companies looking for less expensive ways to implement AI at scale.

According to Artificial Analysis, DeepSeek’s V4-Flash model received a score of 50 out of 100 on its Intelligence Index, which aggregates scores from nine criteria that include workplace-style tasks, reasoning, and coding.