Fable 5 delegates to Sonnet 5: 96% performance at 46% cost

Anthropic is advising developers to stop running Claude Fable 5 as a lone executor and instead use it as a manager that delegates work to Sonnet 5, cutting inference cost while retaining most of Fable 5's performance. The company published two patterns — an Advisor pattern and an Orchestrator pattern — and published benchmark results for each.

How do the Advisor and Orchestrator patterns work?

The Advisor pattern runs Sonnet 5 as the executor and only calls Fable 5 when Sonnet needs guidance, while the Orchestrator pattern uses Fable 5 as a planner that distributes tasks to multiple Sonnet 5 worker agents. Both patterns run through Claude Managed Agents, and each sub-agent uses its own cache to avoid duplicate context costs. The documentation notes the Advisor pattern typically causes Fable 5 to be called roughly once per task.

In practice the Advisor pattern keeps the bulk of execution on Sonnet 5 and consults Fable 5 sparingly. The Orchestrator pattern places Fable 5 at the planning layer and pushes execution to Sonnet 5 workers, which can run in parallel and reuse cached context per sub-agent.

How much performance and cost change did Anthropic measure?

On SWE-bench Pro, Anthropic says the Sonnet 5 executor plus Fable 5 advisor combo reaches about 92 percent of Fable 5's solo performance while costing 63 percent as much. On BrowseComp, using Fable 5 as a planner that delegates to Sonnet 5 workers delivers 96 percent of Fable 5's performance at 46 percent of the cost. Those are the specific, source-published data points Anthropic shared for the two patterns.

Anthropic highlights that the Advisor pattern involves Fable 5 being invoked roughly once per task, which helps explain how the system keeps calls to the expensive model limited and reduces overall token costs. Both patterns use Managed Agents plus per-agent caches to avoid repeated context charges.

Why this change is happening

Anthropic frames these tactics against growing price pressure in the market. The piece notes Chinese open-source models are already undercutting Western pricing, and it points to GPT-5.6 Sol as "much cheaper per token" and reportedly more token-efficient. Turning Fable 5 into a planner or advisor is presented as a way to preserve the model's capabilities where they matter while shifting most token-heavy execution to Sonnet 5.

The company is effectively trading full-model accuracy for a cheaper, hybrid routing: keep the expensive model for strategy and edge cases, use a smaller model for steady-state execution. The published benchmarks show that this tradeoff can recoup a substantial share of raw performance while slashing cost.

What to watch

Look for two concrete signals. First, whether third-party benchmarks reproduce the 92 percent and 96 percent figures Anthropic published for SWE-bench Pro and BrowseComp. Second, watch for how teams actually implement Claude Managed Agents with per-agent caching in production, since the cost wins hinge on avoiding duplicate context costs and limiting Fable 5 calls.

If Chinese open-source models or GPT-5.6 Sol continue to push token prices down, expect more architectures that follow this manager-then-executor pattern. Anthropic’s documentation contains the deployment details for both patterns.

Bottom line

Anthropic is steering users to treat Claude Fable 5 as a higher-level planner or advisor and Sonnet 5 as the workhorse executor. The company’s published numbers show sizable cost reductions — to 63 percent and 46 percent of previous inference cost in two benchmarked patterns — while retaining most of Fable 5’s performance.