1 min read
Add as a preferred source on Google

Arena's Agent Arena Lists Nearly 3 Million Sessions Across 65 Models

Claude Fable 5.2 (Max) led the Nov. 13 snapshot with a Net Improvement score of 28.61% ± 3.72%.

Metal pegs rise across a sculpted leaderboard in staggered rows / TokenPost.ai
Metal pegs rise across a sculpted leaderboard in staggered rows / TokenPost.ai

Arena’s Agent Arena board listed 2,979,271 sessions across 65 models in its Nov. 13 snapshot, expanding the scale of its real-interaction comparison of AI agents.

Claude Fable 5.2 (Max) ranked first, with a Net Improvement score of 28.61% ± 3.72%. The board offers Overall, Code, Chat and Work views.

Agent Arena evaluates real Agent Mode interactions using causal tracing. Its measures include confirmed success, praise versus complaints, steerability, Bash reliability, tool hallucination and Net Improvement. Evaluation signals also include file downloads, disapproval events and retries.

Arena launched the leaderboard June 4, 2026. The service assesses models on tasks carried out in Agent Mode, rather than only on isolated answers. TokenPost previously covered Agent Arena’s rankings and cost-performance view.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…