Google DeepMind and Xiaomi gain ground on Agent Arena benchmark leaderboard

Google DeepMind and Xiaomi have gained ground on the Agent Arena evaluation benchmark, Crypto Briefing reported. Their models have secured frontier positions on the leaderboard for complex, high-cost agentic tasks. At the same time, OpenAI and Anthropic are facing cost-performance pressure as rival models improve. [1]

Crypto Briefing reported that Google DeepMind's Gemini 4 Carbon and Xiaomi's MiMo V2.6 Pro reached top positions for the most expensive workloads. Agent Arena is operated by arena.ai to evaluate how artificial intelligence models execute multi-step agent actions. On these tasks, models consume substantial resources, which increases operational costs for enterprise users. [1]

The benchmark results highlight an industry focus on reducing deployment costs for software agents. As FTMQ SI reported earlier, Google DeepMind previously introduced its Gemini 4 Argon model for software engineering and cybersecurity defense. Meta also released its Muse Wonder and Muse Wonder Full models to run multi-step personal and professional agents, according to earlier FTMQ SI coverage. [2][3]

Crypto Briefing reported that high expenses on complex tasks make cheaper models that reliably complete work more appealing to users. The shift puts financial pressure on traditional market leaders OpenAI and Anthropic as competitors offer viable alternative models. [1]