Anthropic's Fable 5.2 beats Opus on cost but recommends sticking with Opus
Anthropic released Fable 5.2, a model that matches Opus 5.5 on coding benchmarks at half the task cost, yet advises developers to keep using Opus for daily work.
Anthropic has launched Fable 5.2, a new AI model designed for complex software migrations and multi-agent systems. Despite outperforming the flagship Opus 5.5 on key coding benchmarks at roughly half the task cost, the company explicitly recommends that most developers continue using Opus 5.5 for their daily coding and knowledge work.
What happened
The release of Fable 5.2 introduces a paradox in model selection. On the FrontierCode 2.0 benchmark, Fable 5.2 operating at high effort achieved a score of 45.3%, nearly identical to Opus 5.5’s top result of 45.5%. However, Fable completed these tasks at approximately half the cost of Opus. This efficiency gain occurs even though Fable’s base pricing is significantly higher, charging $8 per million input tokens and $40 per million output tokens, compared to Opus’s $4 and $20 respectively.
Anthropic attributes this cost efficiency to Fable’s ability to solve difficult problems with fewer total tokens or more direct paths, reducing the overall computational load despite the higher per-token rate. In an internal experiment involving 1,000 agents formalizing Fermat’s Last Theorem, Fable 5.2 completed the task in 30 hours. In contrast, Opus 5.5 with the same configuration only reached 26% completion after 18 hours and made minimal further progress by the 30-hour mark. This suggests Fable is optimized for long-running, high-complexity research tasks rather than standard development loops.
Despite these metrics, Anthropic maintains that Opus 5.5 remains the "strongest daily driver" for most users. The company positions Fable not as a replacement for general-purpose coding, but as a specialized tool for large-scale architectural changes, multi-agent orchestration, and deep research where conventional models stall.
How it works
Fable 5.2 is designed to function as an orchestrator within a multi-model architecture. Instead of handling every step of a workflow itself, Fable can delegate simpler sub-tasks to cheaper models like Opus 5.5 or Sonnet 5.5. It acts as a high-level planner that analyzes dependencies and maps out changes across services, then assigns implementation details to less expensive models. When those models encounter difficulties, they can escalate the problem back to Fable for deeper reasoning.
This delegation strategy aims to reserve expensive reasoning capabilities for only the most challenging parts of a project. For example, during a major software migration, Fable might plan the overall structure while Opus writes the individual code modules. This approach mirrors infrastructure already adopted by companies like Coinbase, which routes tasks based on complexity and cut its internal AI spending by nearly half despite increased token usage.
However, switching between models introduces a "context-switching tax." Fable 5.2 can read the thinking blocks generated by Opus 5.5, allowing it to understand the reasoning process of the cheaper model. The reverse is not true: Opus cannot read Fable’s thinking blocks. This asymmetry means that passing context back and forth may require additional tokens and careful engineering to ensure the cheaper model retains enough information to proceed without redundant re-computation.
Key details
- Fable 5.2 scored 45.3% on FrontierCode 2.0, nearly matching Opus 5.5’s 45.5%.
- Task costs for Fable were roughly half those of Opus on benchmark tests, despite higher per-token pricing.
- Fable costs $8 per million input tokens and $40 per million output tokens; Opus costs $4 and $20.
- In a 30-hour test, 1,000 Fable agents formalized Fermat’s Last Theorem, while Opus agents reached only 26% completion.
- Fable can read Opus’s thinking blocks, but Opus cannot read Fable’s, creating asymmetric context sharing.
- Anthropic recommends starting with Opus and switching to Fable only when high-effort Opus evaluations fail.
Why it matters
For engineering leads and architects, this release challenges the assumption that lower per-token prices always equate to lower total costs. The data suggests that for complex, multi-step tasks, a more expensive model that solves problems more directly can be cheaper than a cheaper model that requires extensive iteration. This shifts the focus from unit price to total task cost, requiring teams to evaluate models based on end-to-end efficiency rather than just API rates.
The introduction of Fable as an orchestrator also signals a move toward heterogeneous AI systems. Developers will increasingly need to build infrastructure that can route tasks dynamically between models of different capabilities. This requires robust context management strategies to handle the asymmetries in how models share reasoning data. Failure to manage these handoffs efficiently could erase the cost savings gained by delegating to cheaper models.
Furthermore, Anthropic’s recommendation to stick with Opus for daily work, despite Fable’s benchmark superiority, highlights a gap between benchmark performance and practical usability. It suggests that for routine coding, the marginal gains in reasoning depth provided by Fable do not justify the operational complexity of integrating a new model into the daily workflow. Engineers must balance raw capability against the friction of adoption and integration.
What you can do
- Audit your current AI workflows to identify long-running, high-complexity tasks that stall or require excessive iteration.
- Test Fable 5.2 on specific migration or research projects where Opus 5.5 has previously failed to reach completion.
- Calculate total task costs rather than just per-token prices when comparing model efficiency for complex agents.
- Design your agent architecture to allow Fable to delegate simple tasks to Opus or Sonnet, reserving Fable for escalation.
- Implement robust context passing mechanisms to handle the one-way visibility of thinking blocks between Fable and Opus.
- Start with Opus for daily coding and only introduce Fable when specific high-effort requirements are not met.



