Logo of DeepSeek Photo: VCG

Chinese AI start-up DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated AI model promising stronger benchmark performance, faster task completion, and shorter outputs than the previous version while lowering usage costs. 

DeepSeek unveiled V4.1-Flash-1017 on Saturday. The new model is an upgraded version of V4.1-Flash, which debuted last month, and is designed to improve performance in reasoning, coding, and AI agent capabilities.

DeepSeek said that the latest model completes tasks faster and generates shorter responses. By reducing task running time and output length while improving performance, the model helps users work faster and at a lower cost. 

The company said that the model's capabilities in game development and 3D modeling have improved.

DeepSeek said that the new model shows improvements in reasoning, programming, agent tasks and vision benchmarks. The company's own evaluations suggest the model outperformed several leading Western competitors in selected tests.

According to the company, the updated model beat GPT 5.6 Sol. V4.1-Flash-1017 scored 77.0 on the software development benchmark DeepSWE v1.1, compared with 74.0 for Anthropic's Claude Opus 5, and 73.0 for OpenAI's GPT-5.6 Sol. In the AutomationBench workflow automation test, the model scored 60.5, ahead of Opus 5's 50.3.

The company noted that the model's capabilities have been improved through reinforcement learning training based on self-feedback, enabling stronger performance in building complete projects from scratch, including tasks such as game development and 3D modeling.

Alongside the model update, DeepSeek said that its application programming interface (API) now supports new Search and Fetch interfaces, which have been made available to the public at a price of 5 yuan ($0.75) per 1,000 requests.

It has also maintained price competitiveness. The expense per coding task for V4.1 Flash-1017 is about $0.20, far less than one-tenth that of comparable US AI models. 

The fee per 1 million API output tokens for DeepSeek's V4.1 Flash-1017 is $0.59 during off-peak hours and $1.19 at peak. OpenAI's GPT-5.6 Sol is $20, and Anthropic's Claude Opus 5 is $25.

The upgrade brings significant performance improvements, which highlights that Chinese AI start-ups are competing on efficiency, specialization and real-world utility, said a Chinese industry analyst.

In another development, Shanghai-based AI company StartLux secured the top spot on Hugging Face's Decision Index 0.3.2 leaderboard. Its StartLux-Decision-v1.1-27B Instant model achieved a global No.1 score of 69.1, leading a pool of 173 publicly reproducible models.

The combination of stronger coding and reasoning performance, faster task completion, and lower output costs is important because it targets what users actually care about: productivity, affordability, and reliability, Liu Gang, chief economist at the Chinese Institute of New Generation AI Development Strategies, told the Global Times on Sunday.

Chinese AI companies have not only demonstrated their technological capability to continuously iterate and optimize models, but also showcased their ability to drive industry innovation and propose more forward-looking solutions, Liu noted.

Global Times