Mixing Claude models: an Opus lead with Sonnet helpers
aiagents
How a lead model delegating to cheaper helper models can cut cost by about a third, and when the setup makes sense.
Using one model for everything is simple, but not always efficient. A lead-and-helper setup splits the work by difficulty.
How it works
A stronger model, such as Opus 5.5, plans the job and reviews the results. Cheaper models, such as Sonnet 5.5, handle the well-defined pieces. Think of a head chef directing line cooks.
In the guide's example, an Opus 5.5 lead with Sonnet 5.5 helpers cost about a third less with similar results.
Where it fits
- Work that splits into clear, independent subtasks
- Research or document review across many sources
- Pipelines where most steps are routine and a few need judgment
Where it fits less well
- Small one-off tasks where the coordination overhead outweighs savings
- Work where every step depends on deep, shared context
Related cost pattern: low first, redo the misses
Running everything at low thinking and redoing only the failures scored about 97% at 17 cents per task. Running everything on high scored 95% at 29 cents.
What to check before adopting it
- Can you define a clear pass or fail check for each subtask?
- Can you see cost per finished job, not just per request?
- Who reviews the combined output before it is used?
Human review matters, especially with sensitive data. See how we think about responsible AI by design, and for agent-style workflows, explore the AI Agents Store.