Google and Xiaomi Models Pick Up Cost-Performance Spots on Agent Arena

Google DeepMind and Xiaomi models have picked up frontier spots in Agent Arena’s cost-performance view for its most expensive agent tasks, where results are considered alongside the cost of running each model. The comparison highlights a price-based view of agent performance alongside the platform’s overall ranking.

Google DeepMind’s Gemini 4 Carbon (High) entered the overall ranking at No. 2 on Oct. 28 after more than 21,900 sessions. It posted a 16.2% net-improvement score, and its median cost was $1.80 per task. That amount was 72% below the cost of the model in first place.

Xiaomi’s MiMo V2.6 Pro led the open-weight category in mid-October, with a 3.8% net-improvement score and costs that can fall to $0.11 per task. Open-weight models make their trained parameters available, so developers can run them on hardware they control. That gives teams deployment options beyond a closed model available only through an API. MiMo’s category lead applies to open-weight models, while the overall ranking covers all models.

Anthropic’s Claude Opus 5.5 led overall performance at about 20% net improvement as of early November. An overall lead and a position on a cost-performance frontier describe different results. Agent Arena’s frontier tracks the strongest available result at each price; models beyond it can be outperformed on price or results. On the highest-cost tasks, Google and Xiaomi models held frontier spots while OpenAI and Anthropic models slipped on that measure. The comparison shows how a model’s overall standing can differ from its position in a price-based view.

Agent Arena rates models on jobs requiring several connected actions, including work with web search and file systems. Its scores reflect task completion and user feedback, assessing how models perform in the service rather than on a fixed set of questions. Since its June 2026 launch, the platform has gathered data from more than two million user sessions. These tasks involve tools and multiple steps, rather than asking a model to return only a single answer.

A prior snapshot placed OpenAI’s GPT-6.1 Terra eighth overall with a 10.6% net-improvement score.