DeepSeek upgrades AI model: “Performs tasks faster and more cheaply”

[Financial News] Chinese startup DeepSeek said it had increased processing speed and lowered costs by upgrading its existing artificial intelligence (AI) model.
According to the Global Times, Yicai and other outlets on the 19th (local time), DeepSeek said this as it recently unveiled “V4.1-Flash-1017,” a minor upgrade to “V4.1-Flash,” which it announced last month.
The model is said to perform tasks faster, produce shorter answers and offer improved performance. DeepSeek also said that its reasoning and coding capabilities, AI agent functions, and capabilities for game development and 3D modeling had improved.
DeepSeek said its own evaluations showed it outperforming Western competitors in some categories, scoring 77.0 on “DeepSWE v1.1,” a coding capability benchmark, and surpassing Anthropic’s “Claude Opus 4.5” (74.0) and OpenAI’s “GPT-5.6 Sol” (73.0).
DeepSeek also unveiled an application programming interface (API) it developed in-house for “search and content collection,” saying the service could be provided for 5 yuan (about 1,000 won) per 1,000 requests.
It also said the cost per coding task was $0.20 (about 267 won), less than one-tenth the cost of competing U.S. models.
DeepSeek claimed its API output-token fees per 1 million tokens—$0.59 (about 788 won) during off-peak hours and $1.19 (about 1,589 won) during peak hours—were lower than those of Claude Opus 4.5 ($25, or about 33,000 won) and GPT-5.6 Sol ($20, or about 26,000 won).
One Chinese analyst said the upgrade “shows that China’s AI industry is competing on efficiency, specialization and real-world applicability.” DeepSeek’s upgrade, made just a month later, is seen as reflecting intensifying competition between Chinese and U.S. AI models.


[email protected] Lee Seok-woo, International Affairs Specialist Reporter