Agent Arena
An arena-style evaluation platform that assesses the ability of large models to perform tasks as agents.
Timeline (last 90 days)
Moonshot AI's Kimi K3.1 Pro Max ranked 11th on the Agent Arena agent leaderboard, with results based on over 4,000 real task tests by global users and a net improvement of 7.6%.
The Agent Arena feature can have up to five models work on the same task in parallel within the Alibaba environment, each running in an independent worktree, with only the winner's code applied to the main workspace.
Agent Arena is still experimental, and the official documentation contains contradictory statements on 'whether the diff is shown before it is applied'.
Gemini 4 Argon High topped Text Arena, ranked eighth in Agent Arena, with a net improvement score of 7.92% and a cost of approximately $0.62 per task.
All Facts (3)
Moonshot AI's Kimi K3.1 Pro Max ranked 11th on the Agent Arena agent leaderboard, with results based on over 4,000 real task tests by global users and a net improvement of 7.6%.
50%UnverifiedThe Agent Arena feature can have up to five models work on the same task in parallel within the Alibaba environment, each running in an independent worktree, with only the winner's code applied to the main workspace.
50%UnverifiedAgent Arena is still experimental, and the official documentation contains contradictory statements on 'whether the diff is shown before it is applied'.
50%