DeepSeek V4.1 Flash Tops Agent Arena Open-Source Rankings

DeepSeek-V4.1-Flash-20261017 (Max) ranks first among open-source models and No. 11 overall on Agent Arena, adding an agent-performance result to its recent standing on a separate coding board.

The ranking draws on about 7,500 sessions of real use. The model recorded about 5.1% net improvement relative to the average model, with a median cost of about $0.20 per task.

It led open-source entries in four measures: task-success confirmation, positive and negative user feedback, Bash command reliability and steerability. The task-success confirmation measure ranked No. 8 across all models.

DeepSeek recently revised V4.1 Flash with self-feedback reinforcement learning aimed at open-ended coding tasks, including building games and 3D models. The agent result follows an Oct. 18 snapshot of a separate coding board in which the model ranked No. 12 overall and led open-weight entries. It evaluates agents on coding and web development tasks that require several steps and tool use.