Reuters original link.
DeepSeek's new AI model is by far the cheapest of well-known models to run, research firm says
https://www.reuters.com/business/retail-consumer/deepseeks-new-ai-model-is-by-far-cheapest-well-known-models-run-research-firm-2026-08-03/
The following is a partial quote from the Yonhap News translation article.
https://www.yna.co.kr/view/AKR20260803124400009
According to Reuters on the 3rd, DeepSeek's latest version of its flagship model has been shown to have significantly lower operating costs compared to its performance.
According to Artificial Analysis, an AI performance analysis firm, DeepSeek's latest version API model 'V4-Flash', officially released on July 31st, showed that the 'operating cost' was about 1/105 compared to Anthropic's Claude Fable 5 in benchmark tests.
V4-Flash charges $0.14 (about 200 won) per 1 million input tokens and $0.28 (about 400 won) per 1 million output tokens.
The average cost per test is estimated at 3 cents (about 43 won), which is significantly lower than the 86 cents (about 1,230 won) of Kimi K3 developed by Chinese competitor WenShotAI, OpenAI's GPT-5.6 Sol $1.86 (about 2,600 won), and Claude Fable 5's $3.15 (about 4,500 won).
DeepSeek is said to be preparing to submit an IPO application for a listing on the mainland Chinese stock market as early as this year, according to sources who spoke to foreign media. DeepSeek's release of another cost-effective model is seen as an analysis.
However, Reuters pointed out that even if the cost appears low on the surface, if it uses significantly more inference and output tokens than other models in the process of generating answers, the actual input cost may be higher.
Despite the low cost, V4-Flash received a score of 50 on Artificial Analysis's comprehensive performance index, Intelligence Index, which is comparable to Google's Gemini 3.6 Flash.
[The rest of the content is omitted]
Artificial Analysis post
https://x.com/ArtificialAnlys/status/2083306229074739285
