Skip to main content

Claude Opus 5.5 Max Dominates Agent Rankings as Domestic Models Hold Strong

Anthropic's New Model Takes the Agent Crown

The latest Arena AI Large Model Weekly Rankings are out, and the agent category has a new king. Anthropic's Claude Opus 5.5 (Max) didn't just enter the leaderboard—it dominated, landing at number one with a task confirmation success rate of 17.29% and a tool hallucination rate of just 0.37%. That's an ultra-low error rate, signaling a significant leap in reliability for agent-based tasks.

But the story isn't just about one model. Domestic players are quietly making their mark. DeepSeek V4.1 Flash (Max), Kimi K3 (Max), Tencent Hy4preview, and Xiaomi MiMo V2.6 Pro all secured spots in the top tier, proving that Chinese models are holding their ground in the mid-to-upper ranks.

Overall and Code: Google Still Leads, But the Gap Narrows

In the comprehensive ranking, Google's gemini-4-argon-high continues its reign at the top, with Anthropic's main models occupying several of the top four positions. On the domestic side, Moonshot's kimi-k3-max sits at 16th, while Alibaba's qwen3.8-max, Xiaomi's mimo-v2.6-pro, and Zhipu's glm-5.3-max maintain stable placements.

The code ranking tells a similar story: gemini-4-argon-high leads, but Claude Opus 5.5 (High) has surged into the top five. Meanwhile, Moonshot and Xiaomi's code processing capabilities remain at the industry forefront—a reminder that domestic models are not just participating, they're competing.

Niche Battlegrounds: Front-End, Video, and Image Generation

Beyond the headline categories, technical iterations haven't slowed in specialized areas. In front-end development, Claude Opus 5.5 Max continues to lead thanks to its excellent interaction and rendering capabilities. Xiaomi's newly added mimo-v2.6-flash and pro versions both show strong potential.

When it comes to AI video and image generation, Google and OpenAI still occupy some top spots, but domestic teams from ByteDance, Alibaba, and MiniMax are performing stably, demonstrating global competitiveness. The market is clearly watching—and next week could bring even more version updates and product upgrades.

Key Points

  • Claude Opus 5.5 Max tops the agent ranking with a 17.29% task success rate and 0.37% tool hallucination rate.
  • Domestic models like DeepSeek V4.1 Flash, Kimi K3, Tencent Hy4preview, and Xiaomi MiMo V2.6 Pro hold strong mid-to-upper positions.
  • Google's gemini-4-argon-high remains first in both overall and code rankings, while Claude Opus 5.5 (High) enters the code top five.
  • Niche areas such as front-end development, video, and image generation see continued competition from both international and domestic players.
  • Next week may bring further updates as large models deepen their applications across scenarios.