A cheapskate's guide to coding with expensive models

The author explores strategies for reducing the cost of AI coding workflows by using expensive frontier models to delegate straightforward tasks to cheaper subagent models or by utilizing a weaker primary model with an expensive advisor via the Oracle pattern. Testing shows that combining a top model like Opus with cheaper subagents can save roughly sixty percent of spending without sacrificing intelligence, though it doubles task latency. Additionally, choosing the right harness and managing prompt configurations are crucial to prevent token waste through busy-waiting. Open article →