Anthropic: 2 Ways to Use Fable 5 with Sonnet 5 to Cut Costs

Anthropic revealed several of its use cases: which are often used with Fable 5

1. Use Fable 5 as an “advisor” (Advisor)

In this scenario, Sonnet 5 acts as the executor and turns to Fable 5 for recommendations.

Most tokens are billed at the executor’s lower rate.

On the SWE-bench Pro benchmark, the Sonnet 5 + Fable 5 (as Advisor) setup achieves about 92% of Fable 5’s quality, while the cost is roughly 63% of the cost of using Fable 5 alone.

With this approach, Fable 5 is called infrequently (about once per task) to define the strategy, and Sonnet 5 does most of the work.

More details in the documentation.

2. The second approach is to use Fable 5 as an orchestrator

In this case, Fable 5 plans the task execution and distributes it among executors (Sonnet 5).

Most tokens are billed at the executors’ lower rate.

In BrowseComp tests, Claude Managed Agents were used with Fable 5 as the orchestrator and Sonnet 5 sub-agents as executors. This configuration delivers 96% of Fable 5’s performance at 46% of its cost.

Resource-intensive research tasks are handed off to Sonnet 5.

More details in the example: anthropics/claude-cookbooksanthropics/claude-cookbooks/blob/main/managed_agents/CMA_plan_big_execute_small.ipynb

P.S. Claude Managed Agents supports both approaches

With sub-agents you can:

- pass complex tasks up to the Fable 5 level in the Advisor role;
- delegate task execution to Sonnet 5 executors.

Each sub-agent uses its own cache, so on repeated calls you don’t have to pay again for processing the same context in full.