Chinese artificial intelligence (AI) startup DeepSeek's new model has recorded the lowest operating costs among major global AI models. The cost per benchmark test is over 100 times cheaper than Anthropic's 'Claude Payable 5.'
According to Reuters on August 2, AI model evaluation firm Artificial Analysis reported that DeepSeek's 'V4-Flash' had the lowest cost for benchmark tests among major global models.
Artificial Analysis estimated the average cost per test for V4-Flash at 3 cents (approximately 43 won). In comparison, China's Moonshot AI's 'Kimi K3' costs 86 cents (about 1,230 won), OpenAI's 'GPT-5.6 Sol' costs $1.86 (about 2,600 won), and Anthropic's 'Claude Payable 5' costs $3.15 (about 4,500 won).
The usage fee for V4-Flash is $0.14 (about 200 won) per million input tokens and $0.28 (about 400 won) per million output tokens. Tokens are the data units used by AI to process or generate sentences.
DeepSeek officially launched V4-Flash on July 31. The company, which is reportedly preparing for an initial public offering (IPO), is leveraging its ultra-low-cost AI model to regain market leadership.
However, it is challenging to assess actual operating costs based solely on token prices. Reuters noted that even with lower prices, if more inference processes and input/output tokens are required to generate responses, the final costs could increase.
Additionally, V4-Flash's performance did not match some high-performance models. In the 'Intelligence Index' compiled by Artificial Analysis, which aggregates results from nine benchmarks including coding, reasoning, and task-oriented challenges, V4-Flash scored 50 out of 100.
This score is on par with Google's 'Gemini 3.6 Flash' but one point lower than Meta's 'Muse Spark 1.1' and China's Z.AI's 'GLM-5.2.' Moonshot AI's Kimi K3 scored 57, while Anthropic's 'Claude Opus 5,' 'Claude Payable 5,' and OpenAI's GPT-5.6 scored at least nine points higher than V4-Flash.
The index is calculated based on evaluations from nine benchmarks, including coding, reasoning, and task-oriented challenges.
DeepSeek is also preparing a higher-performing model, 'V4-Pro,' but has not disclosed an official release date.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.