Claude Fable 5.2 Tops Agent Arena, 1.6x Ahead of Runner-Up, and the Price Dropped

Arena.ai announced the new Agent Arena rankings on November 11: Claude Fable 5.2 (Max effort) took first place with a net improvement of +30.4%, about 1.6x that of second-place Claude Opus 5.5. The ranking is based on 4.7k real agentic sessions from the global community, and Fable 5.2 also ranks first in all three main categories: Code, Chat, and Work.

Let's add some context first. Agent Arena measures a model's ability to call tools on its own and complete tasks over multiple consecutive steps, not single-turn Q&A. Commenters pointed out that net improvement is weighted by session, and the distance between a 30-point lead and one weighting choice may not be as large as it looks. This metric looks at overall performance, and how the weights are set affects the score gap.

Fable 5.2's changes from Fable 5.1 are straightforward: rank went from 5th to 1st, net improvement from +14.8% to +30.4%; median task cost fell from $7.7 to $7.3, 6% lower; API price fell from $10/$50 per million input/output tokens to $8/$40, 20% lower. Arena also notes that this first place is a result under Max effort, while the Claude app and Claude Code default to Medium. Benchmarking against the default setting would lead to a different conclusion.

Put alongside frontier models: Opus 5.5 has net improvement +19.5%, $5.5 per task; Gemini 4 Carbon (High) +17.0%, $1.5 per task; GPT-6-Astra (Max) +14.5%, $6.0 per task. Fable 5.2 is expensive, but its net improvement is high. One user did the math: paying less than $2 more per task for a +30% score is the first leaderboard number he'd be willing to pay for. Another user cares more about a different cost: in agent scenarios, what really burns money is rerunning after a failure midway; if long context doesn't collapse, actual cost may be lower. These two perspectives are more useful than looking at the ranking alone.

In the signal breakdown, Praise vs Complaint improved +73%, the largest jump relative to the next model's +41%; Steerability +33%, ranking first; Confirmed Success +22%; Bash Reliability +14%, ranking third. Steerability is very practical for multi-step tasks: if a human changes instructions midway, can the model keep up? In the evaluation, it took first place.

Category scores: Code +34%, Work +31%, Chat +27%. Compared with Fable 5.1, Code went from 6th to 1st (+19.8%), Work from 4th to 1st (+15.8%), and Chat from 6th to 1st (+12.0%). These are Arena's net improvement numbers, not raw accuracy.

Anthropic released Fable 5.2 and Mythos 5.2 on November 9. The two use the same underlying model; Fable has cybersecurity and biosecurity classifiers, while Mythos is opened to organizations in cybersecurity, biomedicine, and other fields through a verification program. The official word is that output speed is up to 30% faster, and there is also a Fast mode with token generation up to 2.5x faster at $16/$80 per million input/output tokens.

On coding, Anthropic's data is FrontierCode 2.0: Fable 5.2 at High effort is roughly equivalent to Opus 5.5 at Max effort at about half the cost; at Xhigh effort it reaches 49.1%, the highest among released models. High and Xhigh are reasoning effort levels; higher is slower and more expensive. Ordinary users don't need to treat Max as the default.

The science and engineering cases are the focus of Anthropic's announcement. Fable 5.2 solved two open problems in theoretical physics, both related to black holes and holographic duality. Mythos 5.2 found a method to predict the overall structure of AAV capsids, accurately modeling 28/34 capsids in tests. A team of 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours; under the same setup, Opus 5.5 completed 26% in 18 hours and made almost no further progress by 30 hours.

On robotics, on the RoboDojo simulated dual-arm robot benchmark, Fable 5.2 scored 64.9%, Opus 5.5 39.4%, GPT-6 Astra 29.8%, and Fable 5.1 27.6%. Anthropic also had Fable 5.2 use a single prompt to design and build three electronic devices with circuit boards, 3D-printed enclosures, and software; all three worked without electrical modifications.

There is also a highly demonstrative example. Dan Fein used Fable 5.2 to write code that drew a spider web with 7,330 dewdrops on it, without using an image file. Each dewdrop bends light according to real water refraction, reflecting the entire garden upside down.

The ranking is worth looking at, but don't treat it as your everyday default. Anthropic's own advice is to start most coding and knowledge tasks with Opus 5.5, and switch up to Fable or Mythos when the work becomes larger in scope and more complex. Commenters in the Arena thread also reminded people: today's first place may be pulled down by a new model next week. What really matters is two things: the failure rate on long tasks, and the bill after completing the same work.

Reference links: Arena.ai tweet, Anthropic announcement.