Agent Arena Ranks Claude Fable 5.2 First
Arena's Agent Mode leaderboard, which evaluates models on real tasks rather than isolated answers, lists 65 models across nearly three million sessions.
Updated on Nov. 18, 2026 in Artificial Intelligence

Agent Arena's board listed 2,979,271 sessions across 65 models in its Nov. 13 snapshot. Claude Fable 5.2 (Max) ranked first with a Net Improvement score of 28.61% ± 3.72%.
Why it matters
The leaderboard measures how AI models perform when carrying out tasks in Agent Mode, offering a different view than tests of standalone answers. Readers tracking AI tools can use the ranking to compare models on the kind of work they increasingly perform.
Agent Arena's measures include confirmed success, praise versus complaints, steerability, Bash reliability, tool hallucination and Net Improvement, drawn from signals such as file downloads, disapproval events and retries. The board's underlying methodology for weighting these signals against one another was not detailed in the source.
The players
Arena
Arena is the platform that runs the Agent Arena leaderboard, which compares AI models on real Agent Mode interactions.
Claude Fable 5.2 (Max)
Claude Fable 5.2 (Max) is the AI model that ranked first on the Agent Arena board with a Net Improvement score of 28.61% ± 3.72%.
The details
Agent Arena evaluates real Agent Mode interactions using causal tracing and offers Overall, Code, Chat and Work views of its board. The service assesses models on tasks carried out in Agent Mode rather than only on isolated answers.
Timeline
June 4, 2026: Arena launched the Agent Arena leaderboard.
Nov. 13, 2026: Snapshot listing 2,979,271 sessions across 65 models.
Season Trajectory
Agent Arena extends the tradition of public AI leaderboards by scoring models on multi-step agentic work in live sessions rather than on isolated responses. That shift places the competitive contest among AI developers on sustained task performance, a dimension older static benchmarks did not measure.
For people choosing an AI assistant for multi-step work, the board offers a ranking based on Agent Mode sessions rather than single answers. The ranking reflects one platform's measurements, so readers should weigh it alongside other evaluations before deciding which tool to use.
The takeaway
Agent-mode evaluations based on real sessions can reveal how models perform on sustained work that single-answer tests may miss. Readers comparing AI tools should consider which measures match the tasks they actually need done.
Further reading
Explore more on Artificial Intelligence to see how agentic models are being measured and compared.
Source note: This article includes information reported by TokenPost.


