Claude Fable 5.2 Tops Arena’s New Agent Index by 112 Points

Arena has relaunched its agent leaderboard as the Agent Arena Index, built on over 2.9 million user sessions. Anthropic’s Claude Fable 5.2 (Max) is first at 1,448 points, per Arena’s leaderboard dated 18 November.

That is 112 points ahead of second place, Anthropic’s own Claude Opus 5.5 (Max) at 1,336, by our arithmetic. The index launched the day before Anthropic’s first day of trading on Nasdaq.

Business Pill · Concentration Explained

An evergreen Business Engineer explainer on market concentration. It teaches the general idea only and does not discuss Arena or its rankings.

The key insight: On Arena’s relaunched Agent Arena Index, Claude Fable 5.2 (Max) leads by 112 points and Anthropic holds 5 of the top 9 places, at about 3.5 times Gemini 4 Neon’s cost per task, by our arithmetic.

What Arena Changed

Arena says the core method is unchanged. The main change is the scale: “one index point is 1/1000th of a standard deviation of the signal”, with the average July score set to roughly 1,000.

The index combines four signals: Task Completion, Praise vs Complaint, Steerability and Bash Reliability. Arena retired its Tool Hallucination signal and renamed Confirmed Success as Task Completion.

It sits on top of task-specific indexes for Work, Coding, Chat, Task Outcome, Steering and Tool Use, per Arena’s post.

Claude Fable 5.2 (Max) leads Arena's new Agent Arena Index at 1,448 points, 112 ahead of Opus 5.5 (Max), by ou
Claude Fable 5.2 (Max) leads Arena’s new Agent Arena Index at 1,448 points, 112 ahead of Opus 5.5 (Max), by our arithmetic from Arena’s leaderboard.

What Was Posted on X

TestingCatalog’s daily AI brief for 19 November, which says Arena launched the Agent Arena Index ranking agents on nearly 3M real-world sessions, with Claude Fable 5.2 at #1. The brief says it was composed with Grok; the scores in this piece come from Arena’s own leaderboard.

Who Leads

Anthropic holds 5 of the top 9 places: Fable 5.2 (Max) first, Opus 5.5 (Max) second, Sonnet 5.5 (Max) fifth, Opus 5.5 (High) eighth and Fable 5.1 (Max) ninth, by our count from Arena’s table.

Google’s Gemini 4 Neon (High) is third at 1,333, three points behind Opus 5.5 (Max). Gemini 4 Carbon (High), marked preliminary, is fourth at 1,303.

OpenAI’s best entry is GPT-6.2 Astra (Max) in sixth place at 1,291, which is 157 points behind Fable 5.2, by our arithmetic.

What It Costs

The leader is also the most expensive model in the top 12. Fable 5.2 (Max) costs a median $9.15 per task, against $2.61 for Neon, about 3.5 times as much, by our arithmetic from Arena’s cost column.

Opus 5.5 (Max) costs $5.88 a task. At High effort, Opus 5.5 drops to $1.98 and ranks eighth at 1,273.

Google priced Neon at $3/$15 per million tokens, per its 18 November announcement.

Bar chart: median cost per task on Arena's Agent Arena Index - Claude Fable 5.2 (Max) $9.15, Claude Opus 5.5 (Max) $5.88, Claude Sonnet 5.5 (Max) $3.31, GPT-6.2 Astra (Max) $3.24, Gemini 4 Neon (High) $2.61, Claude Opus 5.5 (High) $1.98, Gemini 4 Carbon (High) $1.35. Source: Arena leaderboard, 18 Nov 2026.
Fable 5.2 (Max) costs a median $9.15 per task on Arena's leaderboard, about 3.5 times third-placed Gemini 4 Neon's $2.61 (our arithmetic).

Why It Matters

On 11 November Arena’s older agent board also ranked Fable 5.2 first. The relaunch keeps Anthropic on top as the company starts trading on Nasdaq today.

Another agent benchmark this week, Ramp’s SWE-Bench v2, found the agent harness adds up to 9.8 points for some coding models.

The new index surfaced on X through TestingCatalog’s daily brief, which said it ranks agents on “nearly 3M real-world sessions”.

The Structural Read

As we read Arena’s table, Anthropic’s lead is at the very top; below it, Google’s best model leads OpenAI’s best by 42 points (1,333 against 1,291, our arithmetic).

As we read the cost column, the top spot is a premium product: buyers pay roughly 3.5 times more per task for the first-placed model.

Arena, introducing the Agent Arena Index

“We’re refreshing our Agent leaderboard to provide clearer scores and more useful benchmarks of agentic performance.”

Three Implications

FOR AI BUYERS Fable 5.2 (Max) leads but costs $9.15 a task against $2.61 for third-placed Neon, per Arena.

FOR MODEL LABS Anthropic holds 5 of the top 9 places on Arena’s new index, by our count.

FOR INVESTORS The index ranks Anthropic first on the eve of its Nasdaq debut; it measures agent use, not revenue.

What Is Not Established

Arena’s leaderboard does not show confidence intervals or per-model session counts, so we cannot say whether close gaps, such as Opus 5.5 against Neon, are meaningful.

Two Gemini models are marked preliminary. Scores reflect Arena’s users and tasks, which may not match any one company’s workload.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Agentic Harness War

A Business Engineer analysis of the fight to own the software layer that runs AI agents.

Read the Map of AI →

The Bottom Line

On Arena’s new Agent Arena Index, Claude Fable 5.2 (Max) leads by 112 points and Anthropic holds 5 of the top 9 places, but it costs about 3.5 times as much per task as third-placed Gemini 4 Neon, by our arithmetic.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Scores, ranks and costs are from Arena’s agent leaderboard dated 18 November, read 19 November; the method is from Arena’s post introducing the index. The gaps, the place count and the cost ratio are our arithmetic. We found the item through TestingCatalog’s brief on X, which says it was composed with Grok.

Sources: Arena, ‘Introducing the Agent Arena Index’, 18 Nov 2026 · Arena agent leaderboard, dated 18 Nov 2026, read 19 Nov 2026 · TestingCatalog daily AI brief on X, 19 Nov 2026

Scroll to Top
FourWeekMBA