DeepSeek V4 Pro Stumbles on Arena

Arena is useful for seeing which model people prefer in blind chats, but it is easy to overread as a proxy for raw intelligence. DeepSeek-V4-Pro may still be strong on reasoning and agentic work even if its conversational style or initial vote distribution lands less favorably.

• –Chatbot Arena measures pairwise human preference, so it rewards polish, helpfulness, and taste as much as raw task performance
• –A weaker Arena debut does not negate a model that may still be competitive on coding, math, long context, or tool use
• –Developers should treat Arena as one input alongside task-specific evals, not as the final verdict on a frontier model
• –Early community discussion often swings hard on first impressions, especially before a model accumulates a stable vote history