DeepSeek V4.1 flash-1017 ranks top in open-source model leaderboard

On October 20, 2026, AI Arena announced that DeepSeek-V4.1-Flash-1017 has topped the Agent Arena and Code Arena: WebDev open-source model rankings, securing the 11th position in the overall leaderboard. The model, based on 7,498 real user sessions, achieved a 5.13% overall net improvement, with a median single-task cost of approximately $0.18. It ranked first in task success confirmation, user feedback, and Bash command execution reliability. Task success confirmation was 11.54%, placing it 8th overall, while user feedback and Bash command execution ranked 11th and 12th, respectively. The model also ranked 7th in tool hallucination, with a score of 0.48%. Arena noted that the model is on the Pareto frontier of performance and cost efficiency, meaning it cannot be outperformed in both metrics by other models. DeepSeek’s latest release includes enhanced performance and reduced task time, enabling "wishful programming" with a single command. The company also launched a web search and scraping API, and released a Harness v0.2.1-rc1 update. Additionally, it is reportedly consolidating its Harness, Web, App, and Platform teams into a new Product Department.