DeepSeek V4.1 Flash Tops the Agent Arena Open-Source Leaderboard

DeepSeek-V4.1-Flash-20261017 (Max) has topped the open-source model leaderboard on Arena’s Agent Arena—a leading AI model evaluation platform—and ranks #11 overall. This result is based on real-world user session data from approximately 7,500 interactions: its overall net improvement stands at 5.1% (vs. the average model), and its median cost per task is $0.20. The model leads all open-source models across four core capability metrics—task success confirmation, recognition of user positive/negative feedback, reliability of Bash command execution, and guidance capability. Notably, its “task success confirmation” score ranks #8 overall.