Advisor and orchestrator patterns to cut Claude token cost
Anthropic published two ways to cut Claude cost: a cheap model that asks an expensive model for advice, and an expensive model that hands work to cheap workers. Both work with a prompt.
The main idea
Anthropic shared two patterns that cut what you pay for Claude. If you run Fable or Claude Opus all day, both apply. The expensive model stops doing the work and starts supervising it. It is a config change.
Most people pick the best model and never touch the setting again. An agent spends most turns reading files, running tests and editing three lines. Those turns do not need a frontier model, but you pay frontier rates for each one.
In the advisor pattern, the cheap model runs the loop. When it gets stuck, it calls the expensive model as a tool, asks one question and keeps going. The advisor gets called about once per task, so everything else bills at the cheap rate. In the orchestrator pattern, the expensive model plans and hands the work to cheaper models.
In the video I ran both side by side on a repo where I live code games. The left terminal ran Sonnet and called a read-only Fable subagent. The right terminal ran Fable and spawned worker agents. I used plain prompts. For a reliable setup, use tool calls and defined agents.
Key points
- In the advisor pattern, the cheap model runs the loop and calls the expensive model as a tool, about once per task.
- In the orchestrator pattern, the expensive model plans and cheaper models execute the work.
- Anthropic's figure on SWE-bench Pro for the advisor pattern is 92% of the output at 63% of the cost.
- I tested both with prompts only: Sonnet on the left calling a read-only Fable subagent, Fable on the right spawning workers.
- For a reliable setup, use tool calls and defined agents instead of prompts.
Chapters
- 0:28Why the default model costs money
- 0:50Pattern 1: the advisor
- 1:17Pattern 2: the orchestrator
- 1:56Setting both up side by side
- 3:40Prompts versus defined agents
Save my free seat →