Anthropic recommends using the expensive Claude Fable 5 primarily as a planner that delegates tasks to the smaller Sonnet 5 model. This "Advisor" pattern achieves 92% of Fable 5's solo performance at 63% of the cost, offering a cost-effective alternative for complex AI workflows.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Faced with high operational costs and pressure from open-source models and competitors, Anthropic publicly recommends two "delegation" patterns, letting its flagship model Fable 5 serve as a planner or advisor, while the smaller Sonnet 5 handles the bulk of the work, significantly cutting costs while retaining most performance.
Claude Fable 5 is Anthropic's most powerful model currently, but its high inference cost limits large-scale adoption. Faced with price wars from Chinese open-source models and OpenAI's newly released GPT-5.6 Sol, which offers lower per-token costs and higher token efficiency, Anthropic had to find optimization paths. Recently, Anthropic's developer team publicly recommended two strategies to transform Fable 5 into a "manager" role, letting the slightly less capable but cheaper Sonnet 5 handle most of the execution work.
The core idea of these strategies is to avoid having the top-tier model handle every step; instead, through task decomposition and on-demand consultation, they significantly reduce the number of calls to the expensive model while maintaining high-quality output.
In the Advisor pattern, Sonnet 5 serves as the primary executor, only requesting guidance from Fable 5 when encountering difficult-to-decide problems. Anthropic's tests show that on the SWE-bench Pro benchmark, this combination achieves about 92% of Fable 5's standalone performance, but at only 63% of the cost. Fable 5 is called on average once per task.
This pattern is like having a senior expert intervene only when there's a question, while a skilled assistant handles routine work. It is especially suitable for scenarios where most steps can be automated, but occasional high-difficulty decisions are needed.
The second Orchestrator pattern positions Fable 5 as a planner, responsible for breaking down tasks and delegating them to multiple Sonnet 5 worker agents for parallel execution. On the BrowseComp benchmark, this pattern achieves 96% of Fable 5's standalone performance at just 46% of the cost—almost half.
This pattern's advantage lies in fully utilizing Fable 5's strategic planning capabilities while leveraging parallel processing from multiple Sonnet 5 instances to boost efficiency. Both patterns rely on Claude Managed Agents to manage sub-agents, and each sub-agent maintains its own cache to avoid overhead from reloading context.
The background to Anthropic's strategic adjustment is the escalating price war in global AI models. Chinese open-source models are offering near-parity performance at far lower prices than Western products, while OpenAI's GPT-5.6 Sol is not only more competitive in per-token pricing but also claims higher token efficiency. These external pressures force Anthropic to find economically sustainable ways to keep users willing to use its most advanced models.
Although Anthropic has not explicitly acknowledged price pressure as the main driver, the timing closely aligns with market dynamics. Through this technical share, Anthropic is effectively offering developers a practical cost optimization guide, and may also be paving the way for broader deployment of Fable 5.
The information in this article is primarily sourced from official documentation by Anthropic's developer team, as reported by AI media THE DECODER. Data comes from Anthropic internal testing; actual production performance may vary.
By turning its most powerful model into a 'manager' rather than a do-it-all executor, Anthropic has significantly reduced inference costs with almost no sacrifice in performance. This strategy not only alleviates users' price anxiety but also provides a reference architecture for the economical commercial application of high-end AI models.
Primary report
Primary source