Agent Arena, a platform launched in June 2026 by arena.ai, tracks AI model performance across over 2 million user sessions. The system evaluates how models handle multi-step tasks using real-world tools like web search and file systems, prioritizing live success rates over static benchmarks.
Google DeepMind’s Gemini 4 Carbon (High) reached the second position on the leaderboard on October 28, 2026. The model operates at a median cost of $1.80 per task, which is 72% lower than the current top-ranked model on the platform.
In the open-weight category, Xiaomi’s MiMo V2.6 Pro secured the top spot in mid-October 2026. This model offers costs as low as $0.11 per task and demonstrated a net performance improvement of 3.8%.
Anthropic’s Claude Opus 5.5 maintains the lead in overall performance with a net improvement of approximately 20% as of early November 2026. However, models from OpenAI and Anthropic are facing increased competition regarding the balance between cost and performance.
