Armstrong's cost-cutting and efficiency improvement moves are indeed hardcore. The cache hit rate went from 5% to 60%, AI spending was directly halved, and the token is still rising. Worth studying.

Armstrong shares how to reduce costs and increase efficiency against the backdrop of surging token usage: no limits, controlling costs by optimizing default models, smart routing, and caching strategies. Using open-weight models like GLM 5.2 and Kimi 2.7 to replace expensive general-purpose models, routing matches by task, cache preprocessing and session management reduce token waste, with the hit rate rising from 5% to 60%. As a result, AI spending nearly halved, while token usage continues to grow.