Anthropic's Fable 5.2 Outperforms Opus on Benchmarks, Yet Company Steers Users to Pricier Model

Anthropic unveiled a coding model that outperforms Opus 5.5 on two significant benchmarks and executes particular complex operations at roughly half the expense. Despite this achievement, Anthropic's announcement of Fable 5.2 on Monday positioned Opus as the preferred choice for the majority of developers, describing it as the "strongest daily driver" for coding and knowledge work.

Based on capabilities and pricing, Claude Opus 5.5 is still our recommended model for most everyday work, from coding to knowledge work.

Anthropic

According to Anthropic, Fable serves specialized use cases including large-scale software migrations, multi-agent systems and extended research projects. However, the performance metrics complicate this positioning. The higher-priced model can deliver lower total costs for specific tasks, and Anthropic's suggested multi-model setup introduces additional factors including task delegation, context handling and operational expenses.

When pricier runs cheaper

On the FrontierCode 2.0 benchmark, Fable 5.2 at High effort achieved 45.3%, nearly equaling Opus 5.5's top performance of 45.5%, while completing those same benchmark tasks at approximately half the expense, per Anthropic's data.

Fable carries double the per-token pricing—$8 per million input tokens and $40 per million output tokens versus Opus's $4 and $20—yet its measured task cost remained lower. This disparity illustrates that token-level pricing alone does not determine overall cost efficiency.

Anthropic also shared results from an internal test where 1,000 Fable 5.2 agents formalized Fermat's Last Theorem over 30 hours. Under identical agent setup, Opus 5.5 completed only 26% of the work in 18 hours and showed minimal advancement by the 30-hour endpoint, according to Anthropic's published findings.

The test illustrates the scope of operations Anthropic envisions for Fable, though it does not clarify whether such gains apply to smaller implementations or typical coding assignments.

Model manager mode

Anthropic is also marketing Fable as an orchestrator for lower-cost models, enabling developers to reserve advanced reasoning for tasks that other models struggle to handle reliably.

Fable 5.2 can act as an orchestrator that delegates simpler tasks to Opus 5.5 or Sonnet 5.5, or as an advisor those models consult on harder problems.

Anthropic

In a large software migration scenario, Fable could examine service interdependencies, design modifications across systems, and assign individual coding tasks to Opus or Sonnet. Tougher technical decisions could escalate back to Fable without requiring the costlier model to handle every component.

Coinbase has already deployed a multi-model system that distributes work according to capability needs and cost, reporting that it reduced internal AI expenses by nearly half while expanding token consumption.

Anthropic intends Fable to manage some of these routing choices independently, though actual savings hinge on how reliably less expensive models complete assigned work without escalation. Repeated model calls, context transfers and unsuccessful attempts can accumulate costs throughout extended agent operations.

The context-switching tax

A further challenge emerges when developers alternate between models during an interaction. Fable 5.2 can process thinking blocks from Opus 5.5, but Opus cannot interpret those from Fable.

This asymmetry may require alternative approaches to transmit pertinent information when constructing systems that shift tasks between models, potentially demanding extra tokens and diminishing the financial advantage of deploying cheaper models.

Switch sooner, save more

Anthropic suggests beginning with Opus and transitioning to Fable when Opus's high-effort assessments fall short of project specifications, offering a capability-driven switching approach that may not reflect the lowest aggregate execution cost.

Anthropic's FrontierCode results demonstrate Fable delivering equivalent performance at roughly half the task cost, suggesting developers might realize savings by switching sooner than Anthropic's guidance indicates.