Agent Arena reveals top AI models for complex tasks as Google and Xiaomi gain ground

Google DeepMind's Gemini 4 Carbon (High) debuted at #2 on the arena.ai-run Agent Arena leaderboard on October 28, 2026, after over 21,900 sessions with a 16.2% net improvement, at a median cost of $1.80 per task — 72% less than the #1 model. Xiaomi's MiMo V2.6 Pro took the #1 spot among open-weight models in mid-October 2026 with a 3.8% net improvement and costs as low as $0.11 per task, while Anthropic's Claude Opus 5.5 still leads overall with around a 20% net improvement as of early November 2026. Agent Arena, which launched in June 2026 and has gathered data from over 2 million real-world user sessions, scores models on multi-step tasks with real tools, and the results show Google DeepMind and Xiaomi gaining frontier cost-performance positions on the priciest agentic tasks while OpenAI and Anthropic slip there.

Agent Arena reveals top AI models for complex tasks as Google and Xiaomi gain ground Google DeepMind and Xiaomi models claim frontier spots on the priciest agentic tasks while OpenAI and Anthropic feel the cost squeeze The most expensive tasks on Agent Arena have new contenders. Google https://cryptobriefing.com/markets/alphabet/ DeepMind and Xiaomi models have taken frontier spots on that end of the leaderboard, while OpenAI https://cryptobriefing.com/markets/openai/ and Anthropic https://cryptobriefing.com/markets/anthropic/ models have lost ground. That matters because these are the jobs where AI agents burn the most money. When the bill gets large, the cheapest model that still finishes the work starts to look very attractive. What Agent Arena actually measures Agent Arena is run by arena.ai, and it judges AI models on how well they behave as agents. Here, “agentic” means a model gets a multi-step job and real tools, such as web search engines and file systems, and has to finish it. The platform launched in June 2026. Since then, it has gathered performance data from over 2 million real-world user sessions. The scoring leans on task success and user feedback rather than static benchmark scores. The goal is to rank models by how they hold up in live use, not by how they perform on a fixed exam they may have studied for. Gemini 4 Carbon makes a cheap, loud entrance The headline act is Google DeepMind’s Gemini 4 Carbon High . It debuted at 2 on the Agent Arena leaderboard on October 28, 2026. It got there after over 21,900 sessions, posting a net improvement of 16.2%. AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. The more interesting number is the price tag. Gemini 4 Carbon runs at a median cost of $1.80 per task, which is 72% less than the 1 model on the board. Xiaomi wins the open-weight bracket Xiaomi’s MiMo V2.6 Pro took the 1 spot in Agent Arena’s open-weight models category in mid-October 2026. Open-weight models publish their trained parameters, so developers can download and run them on their own hardware. That gives teams more control over deployment than a closed model sold only through an API. MiMo V2.6 Pro posted a net improvement of 3.8%. Its cost can drop as low as $0.11 per task. Anthropic still leads, but the leaders are paying for it Anthropic’s Claude Opus 5.5 leads overall performance on Agent Arena, with around a 20% net improvement as of early November 2026. Anthropic and OpenAI both continue to show high raw performance. The pressure is showing up on a different scoreboard: cost versus performance. The most expensive Agent Arena tasks are where that tension plays out. This is where Google DeepMind and Xiaomi have picked up frontier spots, and where OpenAI and Anthropic have slipped. A “frontier” in this sense is a cost-performance line. A model on it offers the best results available at its price point. A model off it can be beaten on either performance or cost by something else. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy https://cryptobriefing.com/editorial-policy/ .