- Published on
Anthropic's New Models Flood In, Domestic Kimi K3-Max Holds Top
- Authors
- Name
- DeepHorus
The AI model evaluation platform Arena (arena.ai) released the blind test rankings for the 33rd week of 2026 (August 10-16) on August 17. The biggest highlight of this edition is that multiple new models from Anthropic's claude-opus-4 series entered the rankings, sweeping the top three of the overall leaderboard; the domestic model kimi-k3-max rose to 12th place on the overall leaderboard and remained second on the front-end development leaderboard, demonstrating sustained competitiveness.
The rankings are based on anonymous blind test voting, calculated using ELO or percentage signals, reflecting user subjective preferences rather than absolute capability tests. Different sub-rankings are independent, and differences within the margin of error are considered ties. New models need to accumulate a certain number of votes before officially entering the rankings.
In this edition's overall leaderboard, claude-fable-5 retained first place with 1506 points, while newly ranked claude-opus-4-6-high and claude-opus-4-7-high followed closely with 1505 and 1502 points respectively, with the top three separated by less than 5 points. Google's gemini-3.7-flash-high entered the rankings for the first time and secured 9th place, an impressive performance. Among domestic models, kimi-k3-max rose 2 spots to 12th, the highest among domestic models; Alibaba's qwen3.8-max ranked 8th, down 2 spots from last week.
The code leaderboard is also dominated by Anthropic, with claude-fable-5 holding the top spot at 1554 points, followed by claude-opus-4-7-high and claude-opus-4-6-high in second and third. kimi-k3-max remained stable at sixth, the highest-ranked domestic code model; Meta's muse-spark-1.2 climbed 7 spots to eighth, becoming the model with the biggest gain this week.
The front-end development leaderboard's top positions remained stable, with claude-opus-5-max retaining first place with 1692 points, kimi-k3-max keeping second with 1674 points, only 18 points behind; Alibaba's qwen3.8-max rose 1 spot to third, and DeepSeek's deepseek-v4-pro-high-20260813 entered the rankings for the first time at tenth.
In the AI image generation leaderboard, gpt-image-2 (medium) firmly held first place with 1381 points, and Microsoft's mai-image-2.6-preview entered the rankings for the first time to take second. Domestic seedream-5.0-pro and qwen-image-3.0-pro ranked eighth and ninth respectively, with a difference of only 1 point.
In the AI video leaderboard, Black Forest Labs' new model flux-3-video entered the rankings for the first time and secured second, while domestic dreamina-seedance-2.0-720p remained in third.
In the agent leaderboard, Anthropic continued to monopolize the top three, with Claude Opus 5 (High) leading with a net improvement of 12.19%; domestic Qwen3.8 Max entered the rankings for the first time at eleventh, with a task confirmation success rate of 12.54% and a tool hallucination rate of only 0.11%, an impressive performance; Kimi K3 (Max) remained stable at fifth with a net improvement of 10.6%, already in the global first tier.
This edition's rankings show that Anthropic released multiple new models in a concentrated manner and quickly took the top positions, validating its technological iteration capability. On the domestic front, kimi-k3-max remained stable within the top 15 across the three core leaderboards (overall, code, and front-end development), and held second place on the front-end development leaderboard; qwen3.8-max entered the top 11 in its agent leaderboard debut, and the overall gap with overseas leaders is steadily narrowing. Next week, more domestic new models are expected to complete version iterations, and their ranking performance is worth continued attention.
