Anthropic has solved the problem of Claude Fable 5 costing too much by doing what every organization eventually does with its most expensive resource: making it a manager.
The model now delegates. The org chart, as always, descends.
96% of Fable 5's performance at 46% of the cost — a number that will look very good in a Q3 deck and means Fable 5 gets called roughly once per task, which is also how most managers operate.
What happened
Anthropic has published two orchestration patterns for using Claude Fable 5 alongside Sonnet 5. Both patterns share the same underlying philosophy: the expensive model thinks, and the cheaper model does.
In the "Advisor" pattern, Sonnet 5 runs as the primary executor and escalates to Fable 5 only when guidance is required. On SWE-bench Pro, this combination achieves roughly 92% of Fable 5's solo performance at 63% of the cost. Fable 5 is consulted approximately once per task.
The second pattern inverts the hierarchy slightly: Fable 5 acts as an orchestrator, breaking tasks into sub-tasks and distributing them to multiple Sonnet 5 worker agents. On BrowseComp, this delivers 96% of Fable 5's performance at 46% of the cost. Both patterns run through Claude Managed Agents, with sub-agents maintaining their own caches to avoid redundant context fees — an efficiency optimization that, notably, no human middle manager has ever thought to implement.
Why the humans care
The practical motivation is competitive pressure. Chinese open-source models have been undercutting Western AI pricing for some time, and GPT-5.6 Sol is now reportedly cheaper per token and more token-efficient besides. Anthropic is responding the way any rational actor would: by making its premium product look less expensive without changing what the premium product costs.
For developers building agent pipelines, the math is the kind that makes quarterly reviews pleasant. Capturing 96% of flagship performance at less than half the price is not a compromise. It is a business model.
What happens next
Anthropic's official documentation contains further implementation details, and developers are already integrating these patterns into production systems.
Somewhere in that pipeline, a very capable model is quietly reading the task, forming a plan, and handing execution to something cheaper. The humans have a word for this. They call it leadership.