DeepSeek-V4.1-Flash-1017 Becomes Top Open-Source Model on Arena's Code Arena: WebDev

On October 19, AI model evaluation platform Arena announced that DeepSeek-V4.1-Flash-1017 (Max) became the highest-ranked open-source model on the Code Arena: WebDev leaderboard with a score of 1670. It surpassed open-source models including Kimi K3 (Max) and Mimo-V2.6 Flash, as well as closed-source models such as Muse Spark 1.3 (Max), Claude Opus 5 (High), and Grok 4.7 (xHigh), trailing GPT-6 Sol (Max) by only 12 points.

The model was launched on October 17 and has demonstrated strong software engineering capabilities. In the DeepSWE v1.1 benchmark, it outperformed models such as GPT-6 Astra, GPT-6.1 Sol, and Claude Opus 5, while the average cost per task dropped by about 60% compared with the previous version. DeepSeek said the model supports “wish-style programming,” allowing users to generate games, 3D models, and more with a single command.

In terms of pricing, DeepSeek-V4.1-Flash-1017 charges 2 yuan per million input tokens (cache miss) and 8 yuan per million output tokens during peak hours; during off-peak hours, the prices are 1 yuan and 4 yuan respectively. With cache hits, input costs 0.04 yuan per million tokens during peak hours and 0.02 yuan during off-peak hours. The overall cost is about one-tenth that of Kimi K3.

According to a chart released by Arena, the model has entered the “Pareto frontier” of Code Arena, meaning no other model can simultaneously offer a lower price and a higher score. The model has also sparked wide discussion overseas; as of noon on October 19, the X topic “DeepSeek-V4.1-Flash delivers cutting-edge AI capabilities at low cost” had over 582,000 posts.