Arena: Claude Fable 5.2 (Max) Takes #1 on Agent Arena With a +30.4% Net Improvement, About 1.6x Runner-Up Opus 5.5, and Leads Code, Work and Chat

Background and Industry Pain Points

Static benchmarks increasingly miss how models behave in real agent workflows, where what matters is whether a task actually gets done, whether the user is happy, and whether the model follows corrections over a long session. Arena (formerly LMArena) built Agent Arena to rank models on real community sessions in Agent Mode. As of November 11 it covers 2.81 million sessions across 59 models. Anthropic shipped Claude Fable 5.2 on November 9 and cut API pricing to $8/$40 per million tokens; two days later it topped this leaderboard.

Architecture Highlights and Mechanisms

Agent Arena does not ask users to vote between two answers. It extracts signals from real sessions: Confirmed Success (the task was confirmed done), Praise vs Complaint (the balance of user feedback), Steerability (whether the model adjusts to user corrections), Bash Reliability (command success, excluding external failures such as network or permission errors) and tool hallucination, then rolls them into a net improvement score. The board also reports median cost, time and output tokens per task, and plots a net improvement vs cost Pareto frontier. Ranks carry confidence intervals, and Gemini 4 Carbon is still marked preliminary.

Benchmarks and Measured Results

Per the Arena leaderboard: Fable 5.2 (Max) +30.35%, Opus 5.5 (Max) +19.52%, Gemini 4 Carbon (High) +16.98%, Claude Sonnet 5.5 (Max) +15.95%, Fable 5.1 (Max) +14.82%, GPT-6 Astra (Max) +14.54%. On signals, Fable 5.2 scores +73.44% on Praise vs Complaint, nearly double Opus 5.5 (+40.32%), and +32.96% on Steerability, though its 14.42% Bash Reliability trails Opus 5.5 at 15.10%. It costs $7.28 per task with about 103.6K output tokens per task; time per task is not yet reported. All figures are Arena's, not Anthropic's own.

Developer Playbook

For long-horizon coding or office agents where getting it right the first time matters most, Fable 5.2 is currently the strongest model in community testing and is available through the Claude API and third-party gateways. For high-volume, budget-sensitive work, look at the Pareto frontier: Gemini 4 Carbon runs about $1.46 per task and GPT-6.1 Sol about $1.08, with a bit over half of Fable 5.2's net improvement. Re-test on your own task set: send hard tasks to Fable 5.2 and route routine subtasks to cheaper models.