▲ DeepSeek
Reuters reported that a latest version among the flagship models of Chinese artificial intelligence (AI) startup DeepSeek boasts overwhelmingly low operating costs relative to its performance.
According to AI performance analytics firm Artificial Analysis, benchmark tests showed that the API model of the latest version "V4-Flash," officially released by DeepSeek on July 31, has an operating cost of about 1/105th compared to Anthropic's Claude Fable 5.
V4-Flash charges $0.14 (approx. 200 won) per 1 million input tokens and $0.28 (approx. 400 won) per 1 million output tokens.
The average cost per test was estimated at 3 cents (approx. 43 won), a remarkably low level compared to Kimi K3 developed by Chinese rival Moonshot AI at 86 cents (approx. 1,230 won), OpenAI's GPT-5.6 Sol at $1.86 (approx. 2,600 won), and Claude Fable 5 at $3.15 (approx. 4,500 won).
As sources told foreign media that DeepSeek is preparing to submit an initial public offering (IPO) application for a listing on the Chinese mainland stock market as early as this year, analysts suggest that DeepSeek has once again put forward its signature strength of cost-efficiency.
However, Reuters pointed out that even if costs appear low on the surface, actual deployment costs could rise if the model requires more steps, such as using significantly more reasoning and output tokens than other models during the response generation process.
Despite its low cost, V4-Flash scored 50 points on Artificial Analysis's comprehensive performance index, the Intelligence Index, putting it on par with Google's Gemini 3.6 Flash.
This is 1 point lower than Meta's Muse Spark 1.1 and Zhipu's (Z.ai) GLM-5.2, respectively.
Kimi K3 scored 57 points.
Anthropic's Claude Opus 5 and Claude Fable 5, and OpenAI's GPT-5.6 scored more than 9 points higher than DeepSeek V4-Flash.
This index comprehensively reflects benchmark results across nine categories, including coding, reasoning, and task-oriented assignments.
DeepSeek is reportedly preparing for the official launch of V4-Pro, an even more powerful model, but the official release date has not been disclosed.
(Photo: Getty Images)
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.