The Arena (arena.ai) platform has released its blind test rankings for the 32nd week of 2026 (August 3 to August 9), featuring several new model entries. Among them, the Chinese-developed Alibaba Qwen3.8-Max stood out by securing the 6th spot in the overall leaderboard, achieving an ELO score of 1497, placing it firmly in the top tier.

Meta's newly launched Muse-Spark-1.2 (xHigh) also claimed the 4th position in the overall rankings, emerging as a notable newcomer from the international camp. This week, multiple sub-categories saw new model debuts, with Chinese models achieving breakthroughs in front-end development, coding, and image generation. Note: This ranking is based on anonymous blind test votes on the Arena platform, calculated via ELO scores (or percentage signals for the agent benchmark) to reflect user preferences rather than absolute capability. Different sub-categories are independent and should not be compared across benchmarks; models with scores within an error margin are considered tied, and new models require a sufficient number of votes to be officially listed.

Where to Begin

The top spots of this week's overall leaderboard remained relatively stable. Anthropic's Claude-Fable-5 continues to hold the top position with an ELO score of 1507, followed by Claude-Opus-4-6-Thinking and Claude-Opus-4-7-Thinking in second and third place, respectively. The score differences among all three are within 10 points, rendering them statistically tied. The highlight of the week centered on two new models breaking into the top tier: Meta's newly released Muse-Spark-1.2 (xHigh) debuted at 4th place with an ELO of 1498, while Alibaba's new Qwen3.8-Max secured 6th place with an ELO of 1497. It shares the same score as the 5th-ranked Claude-Opus-4-6, only trailing due to a wider confidence interval, yet it is also considered statistically tied. The top tier is densely packed, with all top 10 models scoring above 1488, indicating fierce competition. Several Chinese models now occupy the mid-to-upper range of the overall leaderboard, with specific rankings as follows.

Why Focus on Coding Rankings

The coding leaderboard remains dominated by Anthropic's models. Claude-Fable-5 retained its top spot with an ELO of 1553, while Claude-Opus-4-7-Thinking and Claude-Opus-4-6-Thinking took second and third place. Model from Moonshot AI, Kimi-K3-Max, rose from 8th to 6th place this week, increasing its ELO score by 11 points to 1542, breaking into the Top 6 and becoming the highest-ranked Chinese model on the coding leaderboard. Alibaba's Qwen3.8-Max ranked 8th with an ELO of 1532, also entering the top 10. Meta's Muse-Spark-1.2 (xHigh) debuted at 15th place with an ELO of 1526. Multiple Chinese models have entered the top 30 in the coding category, detailed as follows.

Front-End Development Rankings

This week's front-end development leaderboard is still topped by Claude-Opus-5-Max with an ELO of 1686. Moonshot AI's Kimi-K3-Max holds steady at 2nd place with an ELO of 1675, a mere 1-point drop, keeping close to the leading international model. Alibaba's Qwen3.8-Max took 4th place with an ELO of 1667, directly entering the top five in a standout performance. Model from DeepSeek, DeepSeek-V4-Flash-High, also debuted at 8th place with an ELO of 1581. Chinese models now occupy multiple positions in the top 10 of the front-end development leaderboard, demonstrating continuous improvement in overall competitiveness. The specific rankings for Chinese models are as follows.

Agent Benchmark Changes

The top of the agent benchmark saw a change this week. Anthropic's Claude Opus 5 (High) rose from 3rd to 1st place with a net improvement rate of 11.99%, while the previous leader Claude Fable 5 (High) dropped to 2nd place with a net improvement rate of 11.66%. The difference between their rates is less than the sum of their confidence intervals, so they are considered tied. Moonshot AI's Kimi K3 (Max) maintained its 5th place with a net improvement rate of 10.08% and a task confirmation success rate of 14.97%, establishing itself in the first tier of the agent benchmark. DeepSeek V4 Flash (High) (20260731) from DeepSeek debuted at 21st place this week with a net improvement rate of 2.84% and a task confirmation success rate of 8.53%, meeting expectations for its debut. Several Chinese models have entered the top 30 in the agent benchmark, with details as follows.

AI Image Generation Rankings

OpenAI's gpt-image-2 (medium) continues to top the AI image generation leaderboard this week with an ELO of 1380. SpaceXAI's newly launched grok-imagine-image-2.0 (low) debuted impressively at 2nd place with an ELO of 1320. On the Chinese side, Alibaba's qwen-image-3.0-pro debuted at 7th place with an ELO of 1258, while ByteDance's seedream-5.0-pro ranked 8th with an ELO of 1257. Both Chinese models entered the top 10, placing them in the first tier. Tencent's hunyuan-image-3.0 ranked 29th with an ELO of 1151, showing stable performance.

AI Video Generation Rankings

The top of the AI video leaderboard remains stable. Google's gemini-omni-flash still holds the top spot with an ELO of 1513, while ByteDance's dreamina-seedance-2.0-720p stays in 2nd place with an ELO of 1479, a gap of only 34 points, nearing the top level. MiniMax's new minimax-h3 debuted at 4th place with an ELO of 1455, directly entering the top five and becoming the latest highlight among Chinese AI video models. Alibaba's happyhorse-1.0 ranked 5th with an ELO of 1428, also maintaining its position. Chinese models now occupy three of the top five spots on this leaderboard, showing a clear advantage.

This week's Arena leaderboard welcomed several significant new models. Chinese large language models have achieved breakthroughs across multiple dimensions, including overall performance, coding, front-end development, image generation, and video generation. The strong debut of Alibaba's Qwen3.8-Max into the top tier demonstrates that the comprehensive capabilities of Chinese large language models are now approaching international first-tier levels. In several specialized sub-categories, Chinese models have already taken leading positions. More new models are expected to complete testing and enter the rankings next week, and we will continue to monitor the changes.