Agent Arena ranked DeepSeek’s updated V4.1 Flash first among models in its Open Source category on Oct. 19, placing it 11th overall in a ranking built from user sessions.
Listed as DeepSeek-V4.1-Flash-20261017 (Max), the model recorded a 5.13% net improvement score relative to the average model across 7,498 sessions. Arena reported a median cost of $0.19 per task. Its ranking draws on real Agent Mode sessions involving tasks that may require tools; Arena said 75.6% of sessions in a recent sample used at least one tool.
The earlier Sept. 10 version ranked 20th overall, with a 2.56% net improvement score and a median task cost of $0.14. The comparison gives the updated model a higher score and overall rank, alongside a higher measured cost per task.
Within Arena’s Open Source category, the updated model led measures of confirmed task success, user praise relative to complaints and Bash command reliability. It ranked second for steerability, which measures how much correction users must provide. That measure also accounts for turns needing no correction and gives greater weight to repeated corrections.
DeepSeek linked the update’s project-building capabilities, including games and 3D modeling, to reinforcement learning with self-feedback. The company made the Oct. 17 update available through its website, app and API, and released the model weights.
Arena’s Open Source label describes its leaderboard category. Released weights alone do not meet the broader definition of open-source AI described by the Open Source Initiative, which also calls for training information, code and rights to use, study, modify and share the system.
