Agent Arena, a leaderboard run by arena.ai, evaluates AI models on multi-step tasks that use tools such as web search engines and file systems. Launched in June 2026, it relies on results from more than 2 million real-world user sessions rather than static benchmark tests. Google DeepMind entered second place with Gemini 4 Carbon (High), after 21,900 sessions and a 16.2% net improvement, at a median cost of $1.80 per task, 72% less than the model ranked first. Xiaomi took the top spot in the open-weight category with MiMo V2.6 Pro, which recorded a 3.8% net improvement and can cost as little as $0.11 per task. Anthropic’s Claude Opus 5.5 retains the highest overall performance, with a net improvement of about 20%, but OpenAI and Anthropic face increasing pressure on the cost-versus-performance ranking.
Source: Read the original article
