A new edition of the Arena weekly large model leaderboard shows the agent track as the focus of this round's competition. Anthropic's newly released Claude Opus 5.5 Max took first place in the agent ranking on its debut, with a measured task-confirmation success rate of 17.29% and a tool hallucination rate as low as 0.37%, showing notable stability when executing complex tasks.
Chinese models performed steadily, with DeepSeek V4.1 Flash (Max), Moonshot Kimi K3 (Max), Tencent Hy4 preview and Xiaomi MiMo V2.6 Pro all reaching the upper-middle tier of the agent ranking. In the overall ranking, Google's Gemini 4 Argon High retained the top spot with several Anthropic flagship models close behind. Among Chinese models, Kimi K3 Max ranked 16th, while Alibaba's Qwen 3.8 Max, Xiaomi's MiMo V2.6 Pro and Zhipu's GLM 5.3 Max held stable positions. The coding ranking was led by Gemini 4 Argon High, with Claude Opus 5.5 (High) improving markedly.
With model deployment scenarios continuing to widen, more model versions are expected in the coming weeks. In areas such as image and video generation, Google and OpenAI occupy some leading positions while ByteDance, Alibaba and MiniMax maintain stable performance among Chinese models.
Source: AITOP100
