A study found switching models mid-session erases prompt-cache savings, costing 5x at 10K tokens and 11x at 60K tokens.
Continue to AI University →